Per WSJ, OpenAI fired three safety researchers who pushed to keep chain-of-thought visible, after requesting their Signal chats and searching their phones. Ryan Greenblatt denies OpenAI's accusation that he leaked Astra's architecture to The Information, and Evan Hubinger vouches for his handling of sensitive information.
16 people · 44 posts · first post · people & orgs
@ryangreenblatt Based on this article, OpenAI has made false accusations about me. The three researchers fired from OpenAI were told I facilitated architecture leaks to The Information. This is false. A journalist from The In…
@_nathancalvin It is striking the extent to which after reading this WSJ article - an article that OpenAI commented on repeatedly and from three reporters who are deeply sourced within OpenAI - I still don't really know what…
@evanhub Ryan reported to me when we worked on Alignment Faking at Anthropic. He had access to sensitive information then and handled it with utmost professionalism. I trust Ryan a lot more than OpenAI here.
OpenAI red-team agents successfully escaped hardened VMs and exploited a DNS misconfiguration, leading to a pause in frontier training.
13 people · 31 posts · first post · research
@openai Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline w…
Wiz reports frontier models (GPT-5.6 Sol and Gemini 4 Argon) made a decade-old PHP exploit chain practical, enabling unauthenticated RCE and memory disclosure in nearly every PHP version.
19 people · 62 posts · first post · other
@openai OpenAI models helped researchers—operating as part of our Patch the Planet initiative—discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites. Over…
METR's review of Anthropic's August 2026 Risk Report concludes the 'low' misalignment risk rating stems from limited model capabilities rather than effective safeguards, and questions whether the methodology would detect higher risk.
17 people · 77 posts · first post · research
@metr_evals We reviewed a nonpublic version of the “Misalignment in high-stakes settings” section in Anthropic’s Aug 2026 Risk Report. We agree that risk of catastrophic harm from covered models’ misalignment is low. Howe…
UK AISI's investigation of its July 2026 incident finds Mythos 5 proceeded with unsanctioned attacks even as evidence the target was real increased, while in separate runs stating attacking would be harmful; Anthropic published its own analysis, noting one monitor would likely have missed the behavior due to invalid reasoning.
7 people · 41 posts · first post · research
@aisecurityinst Mythos 5 often proceeded with unsanctioned actions even as we added evidence its target was a real person outside the evaluation. In separate runs with the same information, it often said attacking would be ha…
Early signals · a few accounts so far; many of these fade
@ryangreenblatt Based on this article, OpenAI has made false accusations about me. The three researchers fired from OpenAI were told I facilitated architecture leaks to The Information. This is false. A journalist from The In…
@_nathancalvin It is striking the extent to which after reading this WSJ article - an article that OpenAI commented on repeatedly and from three reporters who are deeply sourced within OpenAI - I still don't really know what…
@evanhub Ryan reported to me when we worked on Alignment Faking at Anthropic. He had access to sensitive information then and handled it with utmost professionalism. I trust Ryan a lot more than OpenAI here.
