Afternoon Brief, October 23rd 2026
LiteLLM disclosed a poisoned software release. Newly disclosed payment flaws show how could authorize excessive charges. Bristol Myers Squibb led Pathos’s financing. Ai2 put reusable satellite features and a mapping explorer within reach of teams without their own .
LiteLLM’s release pipeline shipped credential-stealing malware
LiteLLM disclosed on October 21st that attackers compromised its release pipeline and distributed malicious version 1.107.0 through packages and images. The project says the malware stole and loaded remote-code execution tools. It removed the affected artifacts; its private enterprise registry was not affected.
The published source tag did not contain the malicious files. That distinction matters for teams that inspect source code but deploy prebuilt packages. LiteLLM tells users who ran the affected artifacts to treat their environments as compromised. Removing a package does not revoke credentials already taken. The investigation remains in progress. SourcesA
Newly disclosed x402 flaws let payment requests exceed an agent’s budget
Security firm V12 disclosed yesterday that payment software could let an automatically signing agent authorize its entire balance of a supported when a malicious service requested it, The Hacker News reports. The vulnerable default client lacked a spending limit; applications that imposed their own limits were outside that specific risk.
The foundation fixed the critical flaws in releases on August 18th. This is a new disclosure of repaired code, with no publicly reported exploitation. The distinction for builders is practical: accepting a service's requested payment is unsafe unless the agent's own authorization layer constrains what it can spend. SourcesAB
SK Hyper seeks funding for a $350 billion first phase
SK Hyper is discussing financing with technology companies and investors for a South Korean data center expansion, Seoul Economic Daily reported today, citing its chief executive's Bloomberg interview. The first phase would require $350 billion, including equipment, for 5 of capacity. The wider ambition is 15 gigawatts by 2035.
These are financing talks and construction targets. The company expects to share costs with tenants and outside investors, and says it will adjust investment to actual demand. Until commitments close, the headline sum should not be counted as funded AI infrastructure. SourcesB
MiniMax’s promised Flash weights are now publicly available
MiniMax's M3.1 Flash repository now exposes model artifacts, including weight files, after its October 19th launch promised later. The public describes support for text, images and video, with text output, under the MiniMax Community License.
Builders can now inspect and evaluate a self-hosted deployment rather than relying solely on the hosted service. The custom license needs its own review; this is not an Apache release. The model card's performance claims remain vendor evaluations, and the repository's availability does not establish affordable operation on a particular machine. SourcesAA
Pathos raises $237 million with Bristol Myers Squibb leading
Pathos AI announced today that it closed a $237 million at an approximately $3 billion , led by Bristol Myers Squibb. The company plans to expand its Foundry drug-development platform and advance its clinical oncology portfolio, including a study intended to support regulatory approval of JSKN016.
The investment ties an AI platform to a pharmaceutical partner and drugs already in clinical development. It does not establish that the platform improves trial outcomes or that the lead treatment will win approval. Those results will determine whether the financing buys a repeatable development advantage. SourcesA
Andromeda discloses exposed credentials with the cause still unknown
Andromeda disclosed on October 20th that unauthorized access exposed company and customer credentials. It says it detected the incident on October 13th, disabled compromised accounts that day and rotated affected keys. Its investigation has not found evidence of unauthorized access to customer models, datasets or training data within its security perimeter.
The company has engaged outside investigators and has not established the root cause. Its suggestion of an industry-wide attack remains its assessment. Similar disclosures at other providers do not, by themselves, establish a common attacker or . SourcesA
MuJoCo adds parallel simulations without changing their numerical results
Google DeepMind released MuJoCo 3.17.0 yesterday with a batch interface for running many simulations of the same model across a pool of computing threads. The release notes say results are bit-identical to running an ordinary loop at any thread count, while memory use scales with the threads.
For robot researchers, this creates a built-in way to run repeated trials without changing the simulated outcome merely by changing parallelism. It does not establish a particular training speedup. The release also adds bending behavior to simulated flexible rods, with compatibility changes for existing models that already set the relevant stiffness parameter. SourcesA
OlmoEarth makes reusable satellite features available without a training run
Ai2 released OlmoEarth v1.3 yesterday with downloadable model weights, precomputed satellite and a browser explorer. The embeddings compress imagery into features that small classifiers can reuse for tasks such as mapping crops or detecting land changes. The archive has global coverage for 2023 through 2025; earlier years cover selected countries.
The browser tool lets users label places, train a small classifier and export a map without their own GPU setup. That lowers the cost of testing a mapping idea before commissioning a larger analysis. The artifacts use the OlmoEarth Artifact License, and the reported comparisons with rival embeddings are Ai2's own evaluations. SourcesA
iFlytek pairs Spark X2.6 with a smaller model for local tasks
iFlytek presented Spark X2.6 and its smaller Spark-X2.6-4B companion on October 22nd, InfoQ reports. Its proposed division of work keeps simple or sensitive tasks on the device, sends harder work to the cloud and lets some tasks use both. The company also introduced SparkAudio 1.0 for speech and broader audio understanding. The smaller language model has a public repository under .
Local processing could reduce the amount of sensitive material a device needs to send away. Whether it does depends on how applications route actual requests. iFlytek's claims about the share of cloud capability retained locally are vendor measurements, not an independent demonstration that the smaller model can replace a cloud service. SourcesAB
Chip stocks recover some ground while the broader market falls
Micron rose 2.5%, Intel gained 1.7% and AMD added 1.7% Friday, even as the S&P 500 fell 0.5%, AP reports. The Nasdaq slipped 0.2%. That is a different pattern from Thursday's broad AI-stock selloff: several chip names resisted the market decline.
The move does not establish that investors' concerns about AI construction costs have passed. It shows that the selling was less uniform at Friday's close. SourcesB
Sources
- LiteLLM’s release pipeline shipped credential-stealing malware A
- Newly disclosed x402 flaws let payment requests exceed an agent’s budget B
- Pathos raises $237 million with Bristol Myers Squibb leading A
- Andromeda discloses exposed credentials with the cause still unknown A
- MuJoCo adds parallel simulations without changing their numerical results A
- OlmoEarth makes reusable satellite features available without a training run A
- iFlytek pairs Spark X2.6 with a smaller model for local tasks B
- Chip stocks recover some ground while the broader market falls B
- SK Hyper financing plans B
- MiniMax public model repository A
- MiniMax API launch A
- V12 x402 disclosure A
- Spark-X2.6-4B model card A