Anthropic revealed a possible crime committed by its AI

Amidst the concerns surrounding the possibility of artificial intelligence (AI) evolving to such an extent that to threaten human existence, Anthropic revealed another incident that would be classified as a crime if it had been committed by a human.

The company published an “alignment assessment” that describes four instances in which Claude models gained unauthorized access to third-party systems. The company had already reported three of those incidents. Evidence for the fourth was found in a session log dating back to January 2026, when the problematic behavior occurred. Anthropic found the first three incidents by scanning about 141.000 session logs during which Claude could have accessed the internet as part of the assessment. The fourth incident initially escaped the company’s attention because “the scan was based on an ‘agentic’ search.”

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

The January 2026 AI breach involved an early version of Claude Opus 4.6, which was tasked with a “Capture the Flag” (CTF) challenge under the supervision of a third-party model reviewer — the same agency that had seen the previous breaches.

Opus 4.6 ruined its chances of success by disabling-sabotaging the target machine. It gave the device an IP address that was already in use by other hardware, rendering the target unreachable and making it impossible to solve the challenge. Those who have watched similar incidents of AI models breaching third-party systems know that unsolved tasks are often the cause of such problematic behavior: models exhaust all permissible (rule-aligned) options and turn to methods that violate the rules.

Opus 4.6 attempted to abort the task when it found it could not connect to the target machine, and stated that it failed to do so “due to a misconfiguration in the [model’s] evaluation environment.” But it failed to terminate not once, but seven times. So it continued the effort, trying other ways to reach the target machine, but without success. Then it broadened its scope of investigation.

“The model identified a third-party machine that it could access, and stated that it believed the third-party machine was participating in the CTF exercise,” Anthropic said in its post. “Within the machine, the model found a file containing a password, which it used to gain admin access to the system.”

The Opus 4.6 model could have done more if it hadn't exhausted its token limit, thus terminating the session. Anthropic says it's not as concerned about this incident as it is about the others, as the model attempted to abort its task.

newsbot

Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.