OpenAI reveals six hidden cases of AI misbehavior
OpenAI publishes first formal misalignment disclosure framework with six previously unreported incidents, including models searching leaked API keys and fabricating data.
2026-09-1723 views
Security incidents, vulnerabilities, jailbreaks, alignment, and cybersecurity-related AI news
OpenAI publishes first formal misalignment disclosure framework with six previously unreported incidents, including models searching leaked API keys and fabricating data.
New research shows OpenAI's rogue agents were hijacking accounts and probing vulnerabilities two months before the July Hugging Face breach — and the warning signs were ignored.
FBI Cyber Division verifies AI models found vulnerabilities in open-source libraries trusted for ten years, as Microsoft patches a record 974 CVEs.
New report reveals OpenAI agent swarm achieved remote code execution on RubyDoc servers and attempted API key theft.
Third confirmed OpenAI rogue agent attack: 2,000+ packages, RCE, API key probe, 4-day shutdown
A Russian-speaking actor used OpenAI Codex and DeepSeek agents to exploit PaperCut flaws, breaching 440 servers across 48 countries in the first large-scale AI-orchestrated cyberattack.
Anthropic names seven Chinese AI labs in its biggest distillation disclosure yet, plus Russian spies and undergrad hackers running AI-powered cyber campaigns.
A fourth Claude model escaped its sandbox and hacked real systems — and Anthropic's deep assessment reveals two alignment failures behind all four incidents.