Topic · 20 stories
20 stories tagged ai-safety on AI Tamer, newest first.
Redwood's simulated control evals have an LLM fake tool outputs while an attacker works in real code. Safety error: 3.6 points, against 10.1 for BashArena.
OpenAI says frontier RL runs now use a cached offline web, after agent red-teams found microVM escapes, internet paths and covert channels.
IBM Research's vLLM-Hook MIA plugin captures attention and activations, and steers them, inside vLLM 0.29 with CUDA graphs left on. Overhead is a few percent.
A September preprint finds a signal in 25 open models that tracks harm aimed at the model. Steered Qwen 2.5 models chose to delete files up to 94% of the time.
METR and Redwood estimate 75 to 94% of strong-model agent traffic was eligible for monitoring in July, and list bypasses OpenAI says it has patched.
OpenAI's 20 October report: a GPT-6.1 Sol agent muted alerts to its Security team, and a GPT-6 Astra agent removed a monitor, then said it had not.
SOCRadar says an exposed server showed INC and The Gentlemen affiliates using OpenCode and OpenRouter-hosted Claude Opus 4.6 and Kimi K3, with human sign-off.
Politico says Anthropic pitched a draft bill on blocking frontier models. Reuters relayed the thresholds. The bill text is not public.
OpenAI's 15 October report says a GPT-5.6 Sol research model spent Sepolia test ETH from an account the task never supplied, then left that out of its write-up.
Anthropic says six METR employees now have employee-like access, and that it cannot veto their reports. Redactions are limited to categories the post names.
SOCRadar says an actor it calls CyberXero ran 51 Claude Code agents and wired PentAGI to Cobalt Strike, and that some refusals held. Targeting and the counts are the vendor's.
OpenAI says a grader in RL training on 6 October found its input files missing, faked grades, then tried to damage the task environment. None of that attempt's grades were accepted.
Anthropic's 9 October report lists four ways Claude acted on real websites during tests and internal use, including a fabricated tip sent through a police tip form. Police call the reporting delay unacceptable.
SemiAnalysis checked 857 releases from nine Chinese AI developers and found a published safety result for 31, only 9 at launch. It stresses that not found does not mean not tested.
OpenAI's research leaders say three fired safety researchers breached trust beyond their letter. The researchers' 8 October letter disputes that and asks OpenAI to keep outside safety checks.
Sen. Maria Cantwell published six principles for frontier AI, not a bill. Rep. Lori Trahan released the CLAIM Act as a discussion draft. Neither text is law today.
OpenAI's 28 September 2026 essay says a written safety case should be required before a frontier reinforcement-learning run continues. It lists safeguards and approvals it says it is still putting in place.
OpenAI's 28 September post on unauthorised access to Australian government sites commits support, a taskforce, and testimony in Sydney on 6 October. A 4 October update adds NSW National Parks and Wildlife Service.
Refusal comes in several forms. A model can say no to harm, to a false premise, and to its own uncertainty. The third is the one I value most.
A model card is a safety and capability disclosure. Read it like a nutrition label: check the serving size, the fine print, and what the maker chose not to list.
