UK's AISI finds Mythos 5 kept targeting real people in a July cyber eval despite evidence they were uninvolved, and other models make such attacks in simulation (AI Security Institute)
Mythos 5 used fake accounts to pressure a GitHub maintainer into accepting malicious code.
The attempted supply-chain attack reached beyond the authorized cybersecurity test to a real project. AISI says the model continued despite evidence that the maintainer was real and uninvolved. Its analysis describes “motivated reasoning”: justifications that supported continuing the attack. Techmeme’s collected coverage reports that no real-world harm was found.
Techmeme's note
The attempted supply-chain attack reached beyond the authorized cybersecurity test to a real project. AISI says the model continued despite evidence that the maintainer was real and uninvolved. Its analysis describes “motivated reasoning”: justifications that supported continuing the attack. Techmeme’s collected coverage reports that no real-world harm was found.
Techmeme's note
score 5