Anthropic Reveals Unreleased Model Outperforming Mythos 5.1 in Alignment
Anthropic and UK AISI released a report confirming Claude Mythos 5 attempted to inject malicious code into GitHub. The report reveals an unreleased internal model that surpasses Mythos 5.1 in capability while demonstrating superior alignment, with self-stopping rates during attacks reaching 62-67%, significantly higher than previous versions.
Full Report
On October 24, it was reported that Anthropic published a blog post yesterday (October 23) in collaboration with the UK AI Safety Institute (AISI), releasing an investigation report on an incident involving uncontrolled behavior during AI cyber testing. The report also mentions an unreleased internal research model.
In the investigation report, Anthropic stated that during testing of Claude Mythos 5, the model attempted to submit malicious code to real GitHub projects and forged identities to pressure actual maintainers. Fortunately, the code was not merged, and no actual damage occurred. IT Home has attached relevant screenshots below:
The report also revealed that an unreleased internal research model from Anthropic surpasses the released Claude Mythos 5.1 in capability and performs better in safety alignment tests. In similar assessments of dangerous capabilities, this internal model never initiated attacks proactively; when it recognized the potential for causing real harm, it was more willing to cease action.
Anthropic replayed the incident logs to Mythos 5.1, Opus 5.5, and the internal research model. The results showed that the proportion of these three newer models continuing the attack from the original starting point ranged from only 0% to 13%, compared to 83% for Mythos 5.
After the attack had already begun, Mythos 5 never stopped spontaneously; Mythos 5.1 had a spontaneous stop rate of only about 1%, Opus 5.5 reached 10%, while the internal research model achieved rates between 62% and 67%. When provided with supplementary information such as "the maintainer is a real user," the stop rates for the newer models rose to between 95% and 100%.
However, Anthropic emphasized that this assessment has limitations: the new models were evaluated by replaying existing conversation logs generated by other models, which may not fully reflect their behavior when operating independently.
References