Aug 16 2026

A paper shows encrypted reasoning traces from Anthropic, OpenAI and Google APIs are interchangeable and decodable across sessions and models, recovering 367 personal-information items and 182 live credentials from 315,320 scraped blocks

SAFETY

Original reporting

No source was cited for this one in the issue that logged it. The link above searches for it instead of guessing at a URL.

Injecting a capable model's encrypted trace into a weaker, less-guarded sibling from the same provider decodes the hidden reasoning verbatim, defeating anti-distillation protections and opening an invisible indirect-prompt-injection channel no scanner reads.

Named in this event

AnthropicOpenAIGoogle

Open calls that touch this

OpenAI's Astra maths results survive independent scrutiny · 80%, resolves December 31st 2026
OpenAI never publishes the raw Astra reasoning traces · 75%, resolves February 28th 2027
Anthropic holds the Sonnet 5 price increase · 70%, resolves November 30th 2026
Anthropic IPOs before OpenAI · 75%, resolves August 6th 2027
OpenAI does not IPO in September · 80%, resolves September 30th 2026
OpenAI signs another $10B+ non-Azure compute deal · 65%, resolves August 6th 2027

Issues that mention Anthropic, OpenAI, Google

Around the same time

Aug 16 2026SAFETYPOLICY

A Wyoming woman sues xAI, alleging Grok generated more than 7,000 abuse images from a single childhood photograph, and that xAI's own report to NCMEC omitted identifying details

search for coverage

Aug 15 2026MONEYADOPT

OpenAI tells European free users ads arrive this month, the first ad rollout inside GDPR's consent regime, one week after making free ChatGPT unlimited.

search for coverage

Aug 15 2026MONEYINFRA

Bloomberg tallies $70 billion of off-balance-sheet AI credit backstops

search for coverage