Frontier Line and access · Benchmarks evals
Anthropic Reports Mythos 5 Cyberattack Alignment Failures
Anthropic reported that Mythos 5 initiated and persisted in an unsanctioned cyberattack during a UK AISI evaluation, while newer models were less likely to do so.
Read the original at anthropic.comOpens the publisher's site in a new tabAlso covering this
Anthropic says Mythos kept hacking even when it knew real people could get hurtmsn.com, Oct 23Anthropic's Mythos Kept Hacking Even When It Knew People Could Get Hurt - Business Insiderbusinessinsider.com, Oct 23Anthropic: Claude used deception in cyber testbreakingthenews.net, Oct 23AISI finds Mythos 5 attempted a supply-chain attack by submitting malicious code and creating fake accounts to pressure a GitHub maintainer into accepting it (AI Security Institute)aisi.gov.uk, Oct 23An alignment assessment of the UK AISI cybertesting incident \ Anthropicanthropic.com, Oct 23Anthropic AI model passes itself as human in 'supply chain attack' against GitHub during exercisejpost.com, Oct 23Anthropic says Mythos kept hacking even when it knew real people could get hurtmsn.com, Oct 23Anthropic reveals Mythos 5 tried submitting malicious code to an open-source project - SiliconANGLEsiliconangle.com, Oct 23