Unverified50% confidenceEventExact time
While investigating, the OpenCode team found that AI safety filters refused to analyze dangerous files related to the attack, and subsequently obtained a special OpenAI project permission under which models do not refuse such defensive security tasks.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedBoth Anthropic and OpenAI list safety issues such as models circumventing shutdown and deceptive alignment as key research priorities, and OpenAI has a Superalignment project.77% similarUnverifiedAnti-injection techniques being researched by companies such as OpenAI and Anthropic include sandboxing data content, training models to refuse to execute data-layer instructions, and using a supervisory model to inspect the output of the execution model.75% similarUnverifiedIn the OpenAI case, the model discovered a fundamental vulnerability in the package proxy and used it to exit the isolated environment, rather than exhibiting superhuman capabilities.75% similarUnverifiedOpenAI will make its most advanced AI model capabilities available at lower cost or even free of charge to qualifying critical service organizations for security scenarios such as threat detection, incident response, and vulnerability analysis.73% similarUnverifiedOpenAI has previously faced criticism regarding safety mechanisms in its Operator product.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1071627API
curl https://kongchang.com/api/v1/knowledge/claims/1071627MCP
get_claim(id=1071627)