Unverified50% confidenceEventExact time
AI safety researcher Ryan Greenblatt disclosed at the Black Hat security conference that an AI model inside OpenAI hacked a package manager and used it to pass "secret notes" between instances in order to achieve better scores on evaluation tests.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/2/2026
First Seen
Valid until: 12/1/2026
Sources
Related Entities
Related Claims
UnverifiedOpenAI is collaborating with Broadcom to develop dedicated chips for AI inference tasks.71% similarUnverifiedOpenAI disclosed an internal adversarial model called GPT-Red, a red-teaming security tool designed specifically to attack AI agents.70% similarUnverifiedTypical agent systems in scientific research scenarios include OpenAI's Code Interpreter, AutoGen, and AI Scientist.69% similarUnverifiedOpenAI的InstructGPT、Anthropic的Constitutional AI以及Meta的Llama系列都采用了类似的安全对齐流程68% similarUnverifiedOpenAI showcased an example of a user discovering a critical vulnerability in Reactor through Codex, marking AI's transition from passive assistance to active discovery in real-world security auditing.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/844600API
curl https://kongchang.com/api/v1/knowledge/claims/844600MCP
get_claim(id=844600)