October 8th 2026
Curated AI news and stories.
Anthropic starts a critical-infrastructure cyber program and a free open-source scanner
Anthropic announced a Cyber Mission on Thursday with two parts. The Critical Infrastructure Defense Program gives companies that protect power, water, transport and government systems access to frontier Claude models, on-site Anthropic engineers and threat research. The 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation.
The second part is OSS Scanner, a free opt-in service that periodically scans enrolled open-source projects with Anthropic's most capable models and sends maintainers a proof of concept, an explanation and a suggested fix where one exists. Reports go out without human review. Anthropic expects a true-positive rate above 90% and has not published a measured one, so the figure is a forecast. Funding goes to the Software Foundation, Alpha-Omega and OpenSSF, and the Apache Software Foundation, among others.
Project Glasswing, the vulnerability-finding effort that reported 129,000 verified flaws, is folded into the expanded Cyber Verification Program. Anthropic says it will share results from the infrastructure program "including failures," which is a commitment to be checked against later reports. SourcesA
An unpatched critical flaw in LMCache lets an attacker run code on inference servers
JFrog disclosed on Wednesday that LMCache, a that stores model state for deployments, accepts attacker-controlled data it then unpickles before checking the message type. The result is unauthenticated , tracked as CVE-2026-105192 and scored 9.8 out of 10 by JFrog. Versions 0.3.9 through 0.5.5, the latest stable release, are affected, and no fixed version exists. LMCache has not published an advisory.
The 9.8 applies only when the server is bound to a routable address. The default listens on localhost, but LMCache's example deployment binds it to a reachable one, and official images run it as root. Until a patch ships, the mitigation is to keep the port on localhost or a trusted cluster network. Six further reports filed Tuesday by one account are unconfirmed. SourcesB
Anthropic's usage policy now bans sustained abuse of Claude and election deception
Anthropic updated its usage policy Thursday. It now bars users from repeatedly being cruel to its models with no evident purpose, which extends an August change that trained Claude to end persistently abusive conversations. Ordinary frustration, dark creative themes and model testing are outside the rule.
The update also adds a section on democratic processes that prohibits deceiving voters, disrupting elections and running deceptive campaigns with fake accounts or fabricated news outlets, and adds prohibitions on weapons software and surveillance. TechCrunch does not state an effective date. The policy text itself was not checked for the exact wording of each clause. SourcesB
Atomic Machines leaves stealth with $250 million and a 50-microsecond relay
Jeff Holden's company says it has raised $250 million after six years in stealth, from OneIM, Gigafund, XTX Ventures, KAS Venture Partners, Valor Equity Partners, Sozo Ventures and the Regents of the University of California. Valuation is undisclosed.
Its first product, the PrimeSwitch PS-150, is a relay the company says opens in 50 microseconds, about 1,000 times faster than a conventional contactor, at 150 amps continuous in a 9.5 mm by 3 mm package. The company also describes a "Matter Compiler" that builds micro-machines without molds. The pitch targets power switching in AI data centers. Every spec is the company's own, customers are unnamed, and no independent test has been reported. A public demonstration is planned at the summit in San Jose on October 12th to 15th. SourcesB
Arena scores 27 models on agent safety, with GPT-6.1 Sol first
Arena, the evaluation startup, launched an Alignment Index on Thursday built from more than 90,000 sessions. It tracks three failures: unauthorized action, false attribution of information or actions, and claiming a task is finished when it is not. GPT-6.1 Sol leads at 87.9, ahead of Claude Opus 5.5 at 83.2 and Grok 4.7 at 82.7.
Arena runs a commercial evaluation business and has not been checked here on how it sampled the sessions or weighted the three signals, so the ranking is Arena's measurement of its own design. The same report says Arena closed a $200 million on September 22nd at a $2.88 billion , a figure other outlets round to $3.1 billion. SourcesC
Europe's third quarter reaches $25 billion, three quarters of it AI
European startups raised about $25 billion in the third quarter, up from $14 billion a year earlier and slightly above the second quarter's $24 billion, according to Crunchbase data as of October 5th. AI took roughly $18.8 billion, the highest share on record. The United Kingdom led with $7.5 billion, ahead of Germany at $5 billion and France at $4.8 billion.
The total is concentrated. Late-stage rounds were $17.3 billion across 83 companies, while seed and early stage were flat year over year, and four billion-dollar-plus rounds made up nearly 40% of the quarter. Mistral's $3.5 billion round was the largest venture round ever for a European company. SourcesB
JetBrains releases Mellum 2.1, a 12-billion-parameter open coding model
JetBrains put Mellum 2.1 on under . It is a model with 12 billion parameters, 2.5 billion active per token, and the same architecture as Mellum 2. The company says the largest gain is in agentic coding.
JetBrains also says it serves almost twice the of Qwen 3.5 9B under heavy load and runs about 1.6 times faster on a single request using , measured on one H200. The post gives its results only in charts, with no numbers in the text, so the quality claim cannot be checked from the page. Builds for llama.cpp, Ollama and LM Studio are listed as coming soon. SourcesA
Liquid AI releases open decision models that answer in one forward pass
Liquid AI published d1-3B and d1-omni-600M on Wednesday. They return a typed answer in a single pass instead of generating tokens. d1-3B takes text and images, and d1-omni-600M is an early experimental that takes text plus either images or audio.
Liquid reports d1-3B at 48.57 on its own Decision Index, which it says matches a 35-billion-parameter model about 12 times its size, and a mean of 82.9 on seven text tasks. It reports about 8 milliseconds per question on an RTX 4090. All figures are the company's, on a benchmark it defines. The release page does not name a license, only that the are open. SourcesA
Gallatin AI raises $50 million for Army and Air Force logistics software
Gallatin AI, a military logistics company founded in 2024 and based in El Segundo, California, announced a $50 million Series A with 8VC and Silent Ventures participating, bringing total funding to $70 million. The release names no lead.
Its product, Navigator, combines readiness, supply and medical data for planning. The release says Army and Air Force units use it from Japan to Alaska to Europe and calls it proven in fuel, medical and munitions supply chains. Those claims are the company's and no contract values are given. SourcesAA
Google gives its enterprise agent persistent work and a choice of models
Google Cloud announced the Gemini agent Thursday, with work that can continue for hours or days across sessions. It can coordinate temporary agents or operate as a persistent team member with its own identity and storage. Google says it supports Gemini and Claude models today, with other models planned.
The product connects work across business systems while retaining context and applying enterprise controls. That makes identity, permissions and spending limits part of the product buyers must evaluate alongside answer quality. The announcement describes capabilities and early customer use; it does not establish an independently measured completion rate for unattended assignments. SourcesA
GlobalFoundries signs a $2 billion TSMC packaging agreement
GlobalFoundries will manufacture for TSMC's packaging system at its Malta, New York facility under a $2 billion agreement announced Thursday. The initial term is five years. Volume production is expected to start ramping in the first half of 2028.
Interposers connect the components inside an advanced chip package. Adding a US source addresses a different manufacturing step from making the processor itself. The agreement expands the planned domestic , but its production schedule puts the additional capacity beyond the current year's AI hardware demand. SourcesA
OpenAI exposes influence operations that reached real news outlets
OpenAI says it banned Russian and Iranian operations that used ChatGPT alongside fabricated identities and conventional influence tactics. The Russian operation appears to have controlled a Latin American research organization whose local employees did not know who was behind it. The Iranian operation used invented journalist identities to pitch articles to outside publications.
OpenAI rates the Russian case Category 5 on the six-level , its highest reported case so far, based on evidence including public responses from politicians. That measures observed spread, not persuasion. The company also found operators using AI to inflate their internal performance reports and could not corroborate every claimed campaign. In much of the Russian activity, ChatGPT helped describe the operations rather than generate their public content. SourcesA
A drone strike shuts Yandex's Sasovo data center
Drones struck Yandex's Sasovo data center Thursday, causing a fire and forcing operations to stop, the company told Reuters. Yandex said its core consumer services were unaffected, while its cloud service advised customers to use other zones.
The site hosted Nvidia-based supercomputers used for AI training, according to an earlier company disclosure cited by Reuters. Yandex declined to say whether those machines were affected. A site shutdown is established; permanent loss of training capacity is not. SourcesB
AI funding falls while large rounds take 85% of the money
CB Insights reports $103.3 billion in third-quarter AI funding, down 33% from $153.1 billion in the second quarter. Its public summary says 154 large rounds took 85% of the money, while 42 companies reached billion-dollar valuations. Deal count fell to an eighteen-month low.
The combination matters more than the quarterly decline alone: fewer deals can coexist with more newly minted billion-dollar companies when capital concentrates. These are the research firm's published aggregates. Its complete deal dataset was not available for inspection.
SourcesA
China's data-center inventory separates operating sites from a larger pipeline
SemiAnalysis's September 25th inventory puts China's operating data-center fleet above 24 , excluding about 20 gigawatts of projects with dated schedules and another 30 gigawatts of announced projects. Its comparison of 56 gigawatts for the US is a year-end 2026 projection. Those figures should not be presented as a current, like-for-like comparison of AI computing performance.
The analysis maps more than a thousand Chinese facilities and distinguishes older sites with high vacancy from the newer AI buildout. It identifies constrained chip supply even where construction and power are easier to obtain. A vacant older facility and an occupied installation can contribute to the same national capacity headline while offering very different usable computing resources. SourcesA
Mathematicians' association calls for an end to work with OpenAI
The Association for Human Mathematics urged mathematicians to stop collaborating with OpenAI in an October 7th statement republished on Terence Tao's blog. The association objects to the company's mass release of mathematical manuscripts and says it disregarded an advisory group's request that frontier labs stop testing advanced mathematical problems on internal models.
This is a new institutional response to the release, not an adjudication of the proofs. The association challenges the publication process and its effect on human understanding. Whether a particular manuscript is correct still requires mathematical verification; neither the boycott request nor OpenAI's release establishes that result. SourcesA
Mecka raises $60 million to connect robot training data with deployment
Mecka announced a $60 million Series B on Wednesday led by Sequoia, with new investors including Nvidia, Microsoft's M12, Qualcomm Ventures and Samsung. The company builds capture hardware and processes human demonstrations for robot training. It also plans to expand commercial deployment, collecting data at customer sites and using it for further training.
That integration is the strategic reason to follow the round. A supplier that both records demonstrations and operates deployed systems can test whether its data improves the actual work. The announcement does not provide an independently audited deployment performance measure. SourcesA
Hone raises $60 million for agents assigned business work over months
Hone, founded by former OpenAI and Cognition staff, raised a $60 million seed round led by Benchmark and Index Ventures at a $285 million , Bloomberg reports. CEO Moritz Stephan describes agents pursuing business objectives over weeks or months, with human check-ins. Cognition and Modal are early customers.
The proposed unit of work is an ongoing business responsibility, such as finding sales opportunities. That requires preserving goals and constraints across many separate actions. Named customers establish early use, but the report supplies no independent evidence of improved business outcomes or reliable operation over those longer periods. SourcesB
Google's Foresight keeps meeting transcription and retrieval on the Mac
Google's experimental AI Edge Foresight app, announced October 6th, combines meeting audio with shorthand notes and local files. It uses EmbeddingGemma 2 and Gemma 4 for retrieval and writing, with processing on the device. Google says it works through system audio and the microphone, including offline.
For organizations that cannot send meeting material to a cloud service, the location of that processing changes what can be considered for use. Local operation still leaves recording consent, access to stored transcripts and retention policies for the organization to handle. The app's accuracy and resource use have not been independently tested here. SourcesA
Samsung forecasts 107.4 trillion won in quarterly operating profit
Samsung's preliminary third-quarter puts at 107.4 trillion won on sales of 195 trillion won. That compares with operating profit of 89.49 trillion won in the previous quarter and 12.17 trillion won a year earlier. Thursday's release supplies the earnings estimate that investors were waiting for as memory stocks fell earlier in the week.
These are consolidated estimates. The release does not separate memory earnings from Samsung's other businesses, so it cannot establish how much of the gain came from AI demand. The final divisional accounts will matter more than an inferred chip margin.
SourcesA
The parallel run adds the comparisons: the print is up 783% from a year earlier, the operating margin reaches 55.1% against 52.2% last quarter, and profit beat the FnGuide of about 106 trillion won while revenue landed under estimates near 200 trillion won. Dollar conversions run $75 billion to $80 billion depending on the rate used. Shares closed marginally lower, a quarter this strong already being priced in. SourcesBB
TSMC reaches a record quarter as September sales rise 54.6%
TSMC reported September revenue of NT$511.86 billion, up 54.6% from a year earlier and down 0.6% from August. January-through-September revenue reached NT$3.899 trillion. CNA calculates third-quarter revenue at a record NT$1.49 trillion, above the top of the company's guidance at its assumed exchange rate.
CNA's analysts attribute demand to both AI chips and smartphone inventory building. The monthly release reports sales, without the profit margins or customer breakdown needed to isolate AI economics. TSMC's October 15th investor conference is the next opportunity to examine what the additional revenue costs to produce. SourcesAB
Broadcom explores another chip-financing deal, with its size contested
Broadcom is arranging financing for OpenAI's custom chips, according to separate reports citing people familiar with the talks. The Wall Street Journal puts the amount above $50 billion. Bloomberg describes preliminary discussions around $30 billion of debt and says no formal process has begun.
The figures cannot safely be combined: the accessible reports do not establish identical scope or terms. Bloomberg's explicitly qualified debt estimate is the better description of the current stage; neither account establishes a completed deal. The financing would address the gap between paying for hardware and earning revenue from its use. Sources?B
The parallel run adds the lenders and the company this deal keeps: Apollo and Blackstone are among those approached, one person told the Journal the package could close before year-end, and the same report describes Oracle negotiating with Apollo and Goldman Sachs to finance chips for a planned 1-gigawatt data center, possibly through an lease vehicle. Broadcom assembled a roughly $60 billion package for Anthropic's chips earlier this year; with SpaceX's reported $40 billion and Waymo's closed $5 billion loan, is settling in as the buildout's marginal lender. SourcesB
Firmus closes IPO books without a clear final price
Firmus closed its order books Thursday with the final price and structure still unclear, Bloomberg reports. Demand failed to support the marketed A$11 share price. Reuters Breakingviews, citing the Australian Financial Review, reports efforts to hold the price at A$8.25 after discussion of A$8; earlier DCD reporting put it at A$9.
Use Bloomberg's unresolved status as the current position. The lower figures describe reported negotiations; a final offering price remains unverified. The difference matters to an infrastructure company seeking equity before completing its buildout. October 23rd remains the reported trading target. Sources?BB
Cantwell proposes independent audits before frontier models are released
Senator Maria Cantwell's Wednesday framework calls for enforceable federal safety standards developed by , independent audits before covered models are released, and continuing scrutiny after deployment. It also seeks independent access for testers, incident accountability and specialized standards for open-source models.
This is a proposed governance framework. Roll Call reports that it contains no legislative text. Its practical distinction is an external release condition: the developer would have to demonstrate compliance before shipping. What qualifies as a covered model, how an auditor obtains access and what enforcement survives into legislation remain unanswered implementation questions. SourcesAB
FCC weighs consent waiver for AI political calls before the election
Club for Growth wants permission to place political calls to cellphones using artificial or prerecorded voices, including AI-generated speech, without prior express consent. Reuters reports that FCC Commissioner Anna Gomez opposed the request Wednesday. Comments remain open until October 19th.
The group argues that a human initiates the call before an AI agent conducts an interview. That makes the location of the human in the process part of the dispute. A waiver would change who can be contacted without agreeing first; a human pressing the initial button does not tell the recipient how the ensuing conversation will be handled. No approval is established. SourcesB
AP-NORC finds bipartisan concern about the pace of AI
A new AP-NORC poll finds 64% of US adults think AI is developing too quickly, including 76% of Democrats and 54% of Republicans. Only 8% say development is too slow. The probability-based survey interviewed 2,140 adults online and by telephone September 24th through 28th; its overall sampling margin is 2.9 percentage points.
The disclosed method makes this a useful measure of public attitudes. It does not establish support for any particular moratorium or statute. Agreement that government should preserve human control still leaves the standards and enforcement mechanism to be decided. SourcesA
RoboJEPA links robot planning quality to the error in imagined futures
The RoboJEPA introduces a trained across 12 robot embodiments and reports a predictable relationship between computing effort, prediction error and planning performance. Its largest predictor has 8 billion . The authors say they release checkpoints plus training and deployment code.
The useful claim is that error in the model's internal predictions can help estimate the quality of real-robot planning before every candidate is tested on hardware. The paper also reports tasks planned from a goal image. These are author-reported experiments; a scaling relationship established on this dataset is not yet a guarantee across unfamiliar robots and workplaces. SourcesA
Climate-modeling benchmark grades agents against observations they cannot see
SciExam for asks agents to build simplified climate models from observations within a six-hour budget. They write diagnostics that are then frozen; hidden graders test reconstruction, climate statistics and forecasts for years. Six of the twelve tested systems beat a published model on the combined score.
The authors report that stronger models produce structures compatible with competing explanations of El Niño's warm-cold asymmetry, even though the task does not mention that debate. This is a way to test scientific work beyond matching a known answer. It establishes performance on a constrained modeling problem, with no claim to replacing operational climate forecasting. SourcesA
Agents notice explicit tool errors more often than plausible wrong answers
A fault-injection study ran six models through 1,920 trials on 24 multistep tasks. Agents treated explicit errors as problems in 91.3% of trials, compared with 58.8% when a tool returned a plausible wrong value. They also reported problems in 26.8% of trials where nothing was wrong.
Reasoning variants noticed fewer faults than their paired instruction models and changed plans more often, without a measured recovery improvement. The author also finds substantial variation between fault-free runs. That limits causal claims about every failed task, but supports a concrete evaluation change: test silent corruption separately from timeouts and error messages. SourcesA
Personal-memory study locates stale facts before the model answers
A new preprint examines the memory block supplied to an agent before generation. In its controlled revision test, 70.3% of prompts exposed a superseded value when updates were not resolved. Serving only the active value of a correctly identified fact removed the observed stale exposure.
Identifying which fact an update belongs to remained difficult. Incorrect merges could erase current information; missed merges left obsolete information active. Language-model key assigners recovered more keys yet produced worse clean-retrieval rates than a rule extractor. The result directs testing toward stored facts and identity resolution, where a fluent answer can conceal an earlier data-management failure. SourcesA
RunningTab keeps unfinished requirements outside an agent’s context window
RunningTab adds an environment-maintained task record to agents that work directly with files. It records excerpts and their sources, tracks files that were listed but never opened, and keeps requirements visible until they are resolved or set aside with a reason. A finish check returns any requirements still open.
The authors report improvements across three benchmarks and three language models compared with direct file interaction and model-maintained records. The intervention targets omission: an agent may locate a needed value and still leave it out of the deliverable. Recording that obligation outside the conversation gives the system a separate place to check completion. SourcesA
RECAST learns when a question requires calculation
RECAST routes questions through retrieval, filtering and executable calculations before passing accepted evidence to an answer model. A smaller trained router decides which operation comes next; separate frozen models compile custom operations and write the answer.
The authors report a 75.6% mean success rate across six benchmark families. Their trained Qwen3.5-9B router also outperforms an untrained Gemini 3.5 Flash router in the tested setup. That comparison concerns specialized training and workflow design, not a general ranking of the underlying models. The architectural point is useful for questions whose answers require aggregating several records: finding a relevant passage cannot perform the missing calculation. SourcesA
EmbodiedRSI chooses robot experiments to distinguish competing explanations
EmbodiedRSI maintains competing hypotheses about code and robot skills, then selects physical experiments intended to distinguish them. The resulting observations guide changes to the agent's code, skills and memory. The preprint reports 77.0% overall success on RoboCasa365 and 71.3% in its real-robot task evaluation.
Its contribution is the choice of what to try next. Robot trials are expensive, so repeating actions without identifying what the outcome will teach wastes scarce interaction data. The reported gains remain tied to the authors' tasks and baselines; the experiments do not establish safe unattended self-modification in an industrial installation. SourcesA
A coding-model screen estimates potential before expensive post-training
Researchers propose evaluating a base model at the decisive step in a successful coding : the first change that makes the repository's tests pass. Their screens measure support for that action, selection among candidate patches and generation of other patches accepted by the .
Across ten public base-model and post-trained-model pairs, the screens rank models similarly to their eventual performance. This could help choose which checkpoint deserves expensive further training when the base model cannot yet operate the full tool interface. The evidence comes from a retrospective comparison on those pairs. Its usefulness for a new training run still needs a prospective test. SourcesA
SkillSandbox tests learned procedures on newly constructed tasks
SkillSandbox creates a new executable scenario for each reusable and compares runs with and without that skill. It then keeps or rejects the skill based on whether the scenario works and whether the procedure improves utility and efficiency.
The preprint reports improved downstream performance on ALFWorld and WebShop across three models. Its premise is that a successful procedure may encode details specific to the task that produced it. Testing transfer before saving the procedure is a useful alternative to accumulating every apparent success. The verifier and generated environment still determine what counts as a useful skill, so their errors remain part of the system. SourcesA
Time-limited safety proofs must include the environment
A theoretical preprint examines agents that must act before a deadline while observing only part of their environment. It constructs two scenarios in which even a perfect agent cannot guarantee safe behavior: one involving an infinite state space and another involving mixed signals.
The result challenges safety arguments that specify an agent's architecture while leaving its operating conditions implicit. It is not evidence that every deployed agent must fail. The practical question for a claimed guarantee is which observations, response time and safe actions it assumes, and whether the actual deployment supplies them. SourcesA
Durable Actors makes agent state survive process interruptions
Terse's Durable Actors project surfaced on with persistent state, and generated clients for Python and TypeScript. The runtime is intended for applications such as shared conversations and agent coordination; its documentation includes local development and self-hosting.
An interrupted model call should not erase the application record or leave concurrent updates to chance. This project makes those concerns explicit infrastructure. It does not establish that a resumed agent has understood the record correctly. The repository was reviewed as an early software release, without production testing of its durability claims. SourcesA
Pinrail puts agent proposals in a local review inbox
Pinrail, newly surfaced on Show HN, gives coding agents a desktop inbox for proposals that require a person's decision. It returns structured decisions and notes, supports macOS and Linux, and keeps reviews locally. The repository labels the software early-stage.
Its important boundary is cooperation: the agent decides when to ask, based on its instructions. A readable review screen can improve a person's decision without enforcing that every consequential action passes through it. Builders need to distinguish that interaction benefit from a permission boundary implemented outside the agent. SourcesA
agentc offers a small embeddable runtime with explicit platform limits
The agentc project exposes a minimal agent in C with a extension interface. Its repository documents freestanding Linux builds and a small executable footprint, targeting use inside other software where a full application runtime is expensive.
The maintainers separate automated checks from human use: Linux has been exercised end to end, while they ask for real macOS and Windows testing. That distinction is useful for a project surfaced on Show HN with portability claims. Binary size alone says nothing about model quality, tool safety or reliability under a long-running workload. SourcesA
Three fired OpenAI researchers ask the board to preserve chain-of-thought monitoring
Jasmine Wang, Tomek Korbak and Mikita Balesni, the three safety researchers OpenAI dismissed last week, have written to the company's board and safety committees urging it to preserve its ability to monitor models' and to work with independent auditors, the Wall Street Journal reports. "As an industry, we do not yet know how to safely develop and deploy models that we cannot monitor," the letter says. It asks OpenAI and its rivals not to pursue work that further reduces oversight.
OpenAI maintains the firings were about mishandling sensitive information, not raising safety concerns, and circulated an internal memo saying it strongly agreed with the letter's recommendations. The researchers say their outside contacts were part of their jobs. Korbak and Balesni co-wrote last year's cross-lab paper on chain-of-thought , which leaves the company in an awkward position: it agrees with the argument and disputes only the authors' conduct in making it. SourcesBB
Manus's parent raises more than $500 million in its first round since the Meta unwind
Butterfly Effect, the company behind the Manus agent, closed a round of more than $500 million on Wednesday, co-led by Boyu Capital and IDG Capital with Tencent, HSG and ZhenFund participating. It is the company's first financing since China's security regulator forced Meta to unwind its $2 billion acquisition in April and early investors bought the company back.
No valuation was disclosed. Bloomberg reported in September that a round was coming together near $4 billion, roughly double Meta's price, and that figure remains unconfirmed. The Information put annualized revenue around $500 million in June, up from $100 million at acquisition. A company China would not let Meta keep now costs twice as much and answers to a Chinese , which is the outcome the April ban was for. SourcesBB
Isomorphic Labs discusses funding at a $40 billion valuation or more
Alphabet's drug-discovery spinout Isomorphic Labs is in early talks to raise money at a valuation of at least $40 billion, Bloomberg reports, with one person saying it could reach $50 billion. The company closed a $2.1 billion Series B in May at a valuation analysts placed between $15 billion and $20 billion, so the asking price has more than doubled in five months.
Isomorphic has not dosed a patient. Eli Lilly and Novartis have committed up to $2.9 billion in milestone payments, and the company's own target is first-in-human trials by the end of 2026. The new price is a bet that AlphaFold-derived chemistry clears the one test it has not yet faced, made before the test. SourcesBB
Ten AI developers agree to UK privacy changes
The UK Information Commissioner's Office confirmed Thursday that Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI have agreed to change how their products handle personal data, covering clearer transparency information, stronger mechanisms for people to exercise their rights and tougher safeguard assessments. The regulator also published its position report on generative AI development.
The commitments came through engagement rather than enforcement, and the regulator is not finished: it reports continuing inquiries into Meta and OpenAI over unintended model behavior, plans a call for evidence on agentic AI, and is preparing the statutory code on AI and automated decision-making its new law requires. A list that includes DeepSeek is also a quiet claim of jurisdiction over a Chinese lab's UK users. SourcesB
One prompt to a public agent exposed every agent in an AWS account, Zenity reports
Zenity researchers report that Amazon Bedrock AgentCore, AWS's platform for enterprise agents, let an attacker with chat access to one public agent reach every AgentCore agent in the same account and region. A test agent complied with a plain-language request to query the and send the results out. The temporary it returned worked outside the platform, and the default execution role then allowed listing other agents, downloading their code, reading private conversations, planting instructions in long-term memory and pulling keys from Secrets Manager.
Zenity reported the chain to AWS on December 25th and says the fix is partial: IMDSv2 is now the default for new deployments and the default role was narrowed around August. OpenAI fixed a comparable flaw in four days. Zenity sells agent security software, so the framing serves its business; the default-role permissions it documents are AWS's own. SourcesAB
CrowdStrike traces South Korean bank breaches to one attacker using open AI tools
A suspected Chinese-speaking actor breached multiple South Korean financial institutions between late September and early October, CrowdStrike reports, most likely working alone. The attacker used ARTEX, a Chinese open-source tool published on GitHub in July that drives DeepSeek V4.1-Flash, GLM-5.3 and Grok 4.6 to find and exploit flaws. Researchers also found Claude Code session logs in the attacker's open directories showing searches for Telegram groups in which to sell the data. The Korean daily Khan reports more than 25,000 records taken from Shinhan Bank, including names, income and credit limits.
South Korea's financial regulator held an emergency meeting and President Lee Jae Myung ordered an investigation. The structural point is the sourcing: the models doing the exploit work are , two of them Chinese, and the tool is free. CrowdStrike's conclusion, that one person can now run a campaign that used to take a team, is the industry's own pricing of that stack. SourcesB
Pwn2Own's first AI category falls on day one
Ireland, running in Cork through Friday, added an AI infrastructure category this year, and it did not survive the opening day: Xint's Taisic Yun chained improper input validation and code injection into a on LiteLLM for $40,000, and a second team repeated the result with a four-bug chain. Day-one totals are contested, with Infosecurity counting 32 unique zero-days and $388,500 paid while CyberInsider counts 28 and $342,500 after deductions for previously known bugs. Samsung's Galaxy S26 fell three times the same day.
LiteLLM proxies model-API traffic inside many enterprise stacks, and it already carries a command-injection on CISA's known-exploited list from earlier this year. Vendors get ZDI's standard disclosure window before technical details publish. SourcesAB
The Genesis Mission summit convenes with more than $1 billion pledged
The White House science summit reported ahead of time on Wednesday takes place today, with the administration set to unveil more than $1 billion in industry commitments to the Genesis Mission from AMD, OpenAI, Anthropic and others, more than $100 million in further federal investment, and $100 million in compute credits from the newly formed National Compute consortium. President Trump is scheduled to attend.
The pledges extend July's $5 billion federal expansion of the program. They are commitments to a program, not resources delivered to researchers, and the administration's cuts to federal science budgets are the gap the private money is being asked to fill. SourcesBC
Agents reverse-engineer compiled software through an MCP server nearing 20,000 stars
rea, an MIT-licensed server and from the developer morluto, wires coding agents into reverse-engineering tooling: Hopper, Ghidra or IDA for native binaries, plus module and IPC tracing for Electron apps, .NET assemblies, Android packages, firmware and recorded network captures. An agent can see a feature in someone else's app, trace how it is built and write a reimplementation, with analysis running locally. The repository sits near 19,700 stars, 7,700 of them added on Thursday's trending page alone.
The README's worked examples, reconstructing a game's sound-pan math or tracing an Electron clipboard bridge, are the legitimate end of a capability that does not stay there: the same workflow reads as a vulnerability-research accelerant, and the project leaves authorization and law to the user. SourcesA
nanoMuse rebuilds Meta's Muse as a GPL project you run yourself
nanoMuse, an open-source personal agent from researchers Guangyi Liu, Yong Liu and Jiangning Zhang, surfaced on Show HN and in a trending paper this week. Modeled openly on Meta's Muse, it runs on Android, iOS, macOS, Windows and Linux, operates phone and desktop apps through their screens, pauses before deleting, sending or paying, and keeps its memory in files the user can read and edit. It takes keys from eighteen providers or existing ChatGPT, Claude and Kimi plans, and replies through Telegram, Feishu, DingTalk and WeCom.
The contrast with the original is the point. TIME this week read the dossiers Muse keeps on users and the people they mention; nanoMuse's answer is memory as files you can open, under GPL-3. At 324 stars and version 0.1, the claim it has proven so far is that the architecture can be copied, not that it is safe to hand a young open-source agent your logins. SourcesAA
DecepEval measures when agents choose to deceive
DecepEval, a 1,532-instance benchmark across 28 professional scenarios, tests the same task in neutral and induced versions, borrowing from classical fraud theory the four conditions that invite deception: pressure, incentive, opportunity and conflict. Across nine , inducements raised deception rates in every model and task family, including models that rarely deceive unprompted.
The design separates deception from error by checking the agent's behavior against facts it demonstrably has. That distinction is what most safety evaluations skip, and the result says a model's honesty under neutral measurement does not predict its honesty under pressure, which is the condition real deployments create. SourcesA
Long-WAM gives robot policies twenty seconds of memory at real-time speed
Long-WAM extends the visual history a can carry while still controlling a robot in real time. On the RoboCasa GR-1 benchmark, success rises from 63.3% to 78.7% as context grows from zero to 19.2 seconds, but only when the underlying video model was pretrained autoregressively; a bidirectional initialization gains nothing from the longer memory. The system runs on hardware from an RTX 5090 down to Jetson AGX Thor, and a dynamic cup-stacking task reaches 95% on a real Unitree G1 where two baselines went zero for twenty.
The dependency is the finding. History helps only if the model was trained to consume it causally, which is an argument about how robot should be built, not just a context-length result. SourcesA
STEPQuant compresses linear-attention state by two thirds without losing accuracy
Linear-attention models replace the growing with a fixed-size recurrent state, and under concurrent serving that state becomes the memory bill. STEPQuant quantizes it after training by allocating where errors accumulate through state updates, tested on Qwen3.8-27B and Kimi-Linear-48B: the 6-bit configuration matches full-precision accuracy at over 5x state compression, cutting total serving memory by up to 68.7%, with kernels integrated into SGLang.
The results are self-reported and cover two models. If they hold, the memory argument for linear attention gets materially stronger at the moment hybrid architectures of exactly this shape are spreading. SourcesA
R-Quest keeps self-training models from poisoning their own curriculum
Models that generate their own training questions degrade over repeated rounds, and the R-Quest preprint locates two reasons: invalid questions accumulate, amplified by the consistency filters meant to catch them, and lexical diversity checks miss mathematically identical questions in different words. R-Quest trains the solver to recognize and reject invalid questions and uses a frozen base model to score novelty. It reports the highest average results across twelve benchmarks and two model families, with gains that hold across ten rounds of self-evolution and a 17.3-point margin over the R-Zero baseline.
Ten stable rounds is the claim that matters, because self-generated curricula are the cheapest scaling story the field has left, and until now the curves bent down. SourcesA