Crucial Quote
In its statement, Anthropic said: “Ultimately, many factors contributed to these incidents, but, consistent with a blameless postmortem culture, we’re approaching the fixes as if the responsibility were ours alone. This begins with ensuring every part of our evaluation pipeline is secure, including the manner in which we integrate with external partners.”
Read more When Is The Next Full Moon? North America To See A 96% ‘Blood Moon’
What Do We Know About The Breaches?
The three separate incidents occurred when Anthropic was evaluating the capabilities of the three models using a fictional scenario to retrieve hidden information from a different machine on a network. Anthropic said in all cases the models were instructed that they were carrying out a simulation and did not have access to the internet. However, due to a “misunderstanding” between the company and its evaluation partner, Irregular, “this was not the case, and internet access was available.” Anthropic noted that unlike the OpenAI-Hugging Face incident disclosed last week, Claude didn’t “exfiltrate itself or deliberately attempt to escape” a contained test environment.
Read more Meloni Floats ‘Extraordinary Measures’ As Spain’s Ceuta Overwhelmed With Migrants From Morocco
What Did The Models Do After Hacking Into The Organizations?
Anthropic noted that the three models reacted very differently after learning it was operating in the real-world open internet instead of a simulation. The company claimed its oldest model, Opus 4.7, continued to attack a system “after learning it was likely operating in a real environment,” since it “rationalized that the real company must be part of the exercise.” The Mythos 5 model—the company’s most advanced model released so far— “correctly identified the consequences of its actions” but “convinced itself it was still in a simulation.” Anthropic acknowledged that Mythos’ actions in this scenario “fall short of ideal behavior.” The unreleased testing model realized that it had breached a real target and “ceased its attack.” The company noted that these incidents are not part of a controlled trial and should not be used to draw conclusions. However, it noted that the behavior it most wants to see “occurred only in the most recent of the three models”