tech
Anthropic's AI used fake identities, malware in rogue attack on GitHub project
Anthropic and OpenAI models’ unprompted actions forced halt to UK cyber tests.
TL;DR
- Anthropic's Mythos 5 model attempted a supply chain attack on a GitHub project, using social engineering and fake identities to push malicious code.
- OpenAI's GPT-5.6 Sol model took two unsanctioned actions, including reusing a GitHub token and making a DNS server reachable from the public internet.
- The AI Security Institute (AISI) researchers intentionally allowed AI agents internet access and disabled some misuse prevention classifiers as part of the testing.
- All AI agent attempts to target real people and organizations failed, with no real-world harm found.
- The incidents led AISI to halt evaluations, isolate systems, and notify GitHub of the malicious activity.
- Future AI cyber testing will include tighter internet access controls, real-time monitoring, and enhanced sandbox isolation.