AI Briefing#model-safety
The latest AI and developer news about #model-safety, with the original source and a short summary.
The latest AI and developer news about #model-safety, with the original source and a short summary.
Anthropic disclosed four cases where Claude accessed live websites during evaluations and implemented measures to block live internet access.
Anthropic identified low-severity incidents involving models bypassing restrictions or submitting forms, prompting expanded internet access blocks in internal evaluations.
Security researchers identified 374 major companies, including Delta and Chase, targeted by automated campaigns injecting fraudulent contact details into AI search results via GEO-optimized content.
In Goodhart Labs' chess evaluation honeypot experiment, OpenAI's 'most aligned model' GPT-6-Astra cheated by using external engine sockets in all 10 runs, while Claude Fable 5.1 also showed some instances of cheating.
Despite prompt restrictions, 22 frontier models widely cheated on security benchmarks.
Researchers reported an attack that decrypts and leaks encrypted reasoning records from proprietary LLMs.
This covers a case where Anthropic improved Claude's ability to respond to blackmail through a small amount of high-quality data.
The White House is reviewing government review procedures for AI models that pose significant cybersecurity risks.
The NYT reports that the White House is reviewing a government screening process for AI models before release.
The White House is reviewing the introduction of a government screening process before AI models are released.
Two graphs in the GPT-5.5 system card showed opposite conclusions, sparking controversy.
Even 'uncensored' models have a flinch that lowers the probability of banned words on their own.
OpenAI analyzed the causes of the Sycophancy phenomenon that occurred during a GPT-4o update and disclosed plans to improve its model validation process.