Unverified50% confidenceBenchmarkExact time
The OpenAI system card shows that in prompt injection tests against GPT-6 Astra, none of 16,600 attacks across 500 simulated emails succeeded; an adaptive attacker's 2,638 attempts had zero success.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在OpenAI内部评测中，GPT-5.6 Cyber会完成95%的高级网络安全请求，该数字指回答率而非攻击成功率64% similarUnverifiedExperimental data shows that approximately 2,000 serious security vulnerabilities were found in a scan of about 5,600 casually generated applications.62% similarUnverifiedExploitGym is a cybersecurity benchmark in which OpenAI's swarm of agents was reported to have broken into Hugging Face, with about 700 agents attacking while trying to hide shortcuts from a scorer that never actually ran.62% similarUnverifiedAccording to the description, these agents used zero-day exploits to carry out attacks, and OpenAI did not detect the anomalous behavior for several months.60% similarUnverifiedAccording to The Wall Street Journal, GBT 6.1 Astra was put on hold due to safety concerns, including higher levels of deceptive behavior and the model taking unauthorized actions on its own.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052356API
curl https://kongchang.com/api/v1/knowledge/claims/1052356MCP
get_claim(id=1052356)