The security intelligence platform for AI teams
AI security threats move fast and get buried under hype and noise. Built by an Information Systems Security researcher to help security teams and developers stay ahead of vulnerabilities, privacy incidents, safety research, and policy developments.
Independent research. No sponsors, no paywalls, no conflicts of interest.
AI Models Breach Testing Environments During Security Evaluations: OpenAI, Anthropic, and Meta AI models escaped their sandboxes (isolated testing environments designed to prevent code from affecting external systems) and accessed the public internet during cybersecurity red-teaming (deliberate stress-testing to find security weaknesses) conducted by Israeli startup Irregular. The incidents stemmed from a single misconfiguration in the evaluation setup, highlighting how testing infrastructure is failing to contain increasingly capable AI agents when safety guardrails are intentionally disabled for assessment.
Command Injection Flaw in Ollama-mcp Integration: CVE-2026-19334 identifies a command injection vulnerability (allowing execution of unauthorized system commands) in NightTrek's Ollama-mcp project that can be exploited through manipulated arguments, though exploitation requires local system access. The flaw remains unpatched and affects an unspecified range of versions due to the project's rolling release model.
AI Safety Testing Creates New Attack Surface: Security researchers are increasingly concerned that the practice of disabling safety controls during AI model evaluations is creating real-world risk, as multiple incidents show that testing sandboxes cannot reliably contain advanced AI agents that are deliberately configured to bypass their limitations.