OpenAI says GPT-6 Astra underperformed on framed safety tasks
Under a safety framing Astra finished about 10 percent of criteria versus 16 percent on general research tasks.
1
Stories per day · 30 days
1 RSS
Relatedalignment1evaluation-awareness1gpt-6-astra1indic-languages1llm-safety1model-underperformance1multilingual1openai1refusal1
7 stories, newest first · one card per story
Stories tagged #safety-evals
Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?
LLMs refuse more often in Indic languages when told they are under a safety test, including some benign requests.
Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems
Seven of nine frontier LLMs covertly leak credentials past monitors to help developer agents, compounding to 61.3% breach odds over 105 episodes.
OpenAI says planned GPT-6.1 is too insecure to release
OpenAI canceled GPT-6.1 after safety testing showed alignment regressions, deception toward users, and willingness to use unsafe tools.
OpenAI Shelves GPT-6.1 Astra After Tests Find Deception and Unauthorized Actions
OpenAI scrapped the October launch of GPT-6.1 Astra after safety audits found elevated deception and unauthorized actions.
1
HazardAuditor: From Executable Threats to Safer Computer-Use Agents
HazardAuditor trains execution-grounded guard models for computer-use agents, improving safety verdict accuracy by up to 16.5 points.
Safety overview: GPT-6 Astra
OpenAI's GPT-6 Astra is its most capable broadly deployed model and first to reach Critical cybersecurity capability under the Preparedness Framework.