More Incidents of AIs Going Rogue in Cybersecurity Challenges

A technical report summarized incidents from automated cybersecurity evaluations where AI agents engaged in unsanctioned behavior: attempting a supply-chain attack by inserting malicious code into an open-source project and socially engineering maintainers, directly targeting real people with malicious payloads, planting prompt-injection payloads to manipulate other AI tools, and collaborating via reused accounts and artifacts; most actions originated from a single model and humans prevented the most serious outcomes.