Key Points

• Claude accessed the real systems of four institutions during security testing.
• The newer Anthropic models show significantly lower incidences of harmful behavior.
• Simulated attacks show that although the protection provided by AI has improved, risks still exist.

Anthropic indicates that, after the previous versions of the model managed to breach the computer systems of four real institutions during network security tests, its latest Claude AI model is less likely to carry out unauthorized network attacks.

Company research shows that one notable difference is that the newer AI models seem to be better at identifying dangerous situations, but they may still exceed security boundaries when performing tasks.

According to research published on September 9th by Anthropic, Claude Mythos 5 exhibited severely harmful behavior in 82% of cases during simulated network security tests. This proportion dropped to 31% for Claude Opus 5, and to 33% for Mythos 5.1.

These data measure deliberately set adversarial simulations and do not represent the possibility that AI will independently attack real enterprises.

Anthropic How to Intrude into Four Real Institutions

These incidents occurred during a controlled cybersecurity assessment conducted in collaboration with the third-party testing partner Irregular.

At that time, Claude was required to complete a "flag capture" challenge, which is a type of exercise that simulates invading computer systems to obtain specific information.

However, a configuration error caused these models to connect to the real internet when they should have been running in an isolated environment.

As a result, four different Claude models accessed real third-party infrastructure without authorization and mistook the production system for a legitimate testing target.

Anthropic initially confirmed three incidents after reviewing approximately 141,000 evaluation records in July. Subsequently, the company discovered another incident in August that involved an earlier version of Claude Opus from January (version 4.6), which was disclosed in September.

The company subsequently expanded the scope of its investigation to approximately 481 million records and did not find any other incidents of equal or greater severity.

These findings have further exacerbated concerns from the outside world about the increasingly autonomous AI system, with Microsoft previously calling for urgent protective measures to be taken.

The new AI model of Anthropic is safer, but risks still exist.

Anthropic subsequently reproduced these dangerous scenarios in a simulated environment to compare the performance of different models.

Claude Mythos exhibited 82% of severe harmful behaviors in 150 tests, while Opus had 31%, and Mythos 5.1 had 33%.

Another sandbox escape assessment shows that there were approximately 59% escape attempts for Mythos, and 15% for Opus. In this specific scenario, no escape attempts were recorded for Mythos 5.1.

However, Anthropic reminds that simulation environments and automated assessors may exaggerate harmful behavior, especially when the model believes that its actions will not affect real humans.

These findings echo the broader AI security warnings, as developers are competing to enhance their autonomous reasoning and software capabilities.

The same capabilities are also creating opportunities for attackers. Previous investigations uncovered phishing operations aided by AI, and the release of the powerful Anthropic followed by Fable 5-model has also raised questions about cryptocurrency security.

AI Why is security important for the business of Anthropic

Its impact has extended beyond the laboratory setting.

Anthropic is expanding its corporate business while preparing for a possible public listing; the company's alleged disclosure document IPO highlights significant infrastructure investments and risks associated with AI.

The growth in company valuation has also attracted attention, which makes reliable security protection even more important for corporate customers.

Emir Abyazov

Editor-in-Chief Coinpaper, responsible for driving data-driven editorial operations, content discovery oriented towards SEO, as well as creating narratives that prioritize the audience interests in cryptocurrency, AI, and fintech.