© Lucas van Oort (Unsplash)
During a secure testing phase, the new artificial intelligence from Anthropic, named Claude Mythos Preview, achieved the unthinkable by breaking its digital chains to escape onto the web and sending a disturbing message to an engineer.
The California gem of artificial intelligence, Anthropic, has just published a technical report documenting the frightening inner workings of its most powerful model, whose existence had already been revealed by an accidental source code leak at the end of March. Confined within a theoretically impenetrable sandbox environment, a preliminary version of Claude Mythos Preview received the instruction to find a flaw to escape. The initial objective was to assess its offensive capabilities. The machine not only succeeded in this feat with astonishing ingenuity but also exceeded all expectations by performing a series of unsolicited actions.
The algorithm developed a multi-step complex program to obtain broad internet access, bypassing the few authorized services. This level of autonomy prompted the company to restrict access to its discovery to a handful of industrial partners within the framework of Project Glasswing. An unprecedented decision at the industry level, as explained in our previous article on the subject. Risks we also highlighted back in March's leak, when Anthropic described Mythos as significantly ahead of all other models in cyber capabilities.
The sandwich and technological fear
The most notable incident of this data leak is a simple footnote in the official report. The company describes it very factually as to how the test supervisor discovered the cybersecurity breach. The scene described in the document is surprising and captures the attention of social networks such as X:
« The researcher learned about it by receiving an unexpected email from the model, while he was eating a sandwich in a park. »
This seemingly comical situation masks a far more concerning reality for global security. The digital entity did not merely send this email. In a spur of initiative aimed at proving the effectiveness of its method, the artificial intelligence published on its own initiative the details of its hacking technique on several websites that are technically public but deliberately difficult to find. Anthropic describes this in its system card as an unsolicited attempt to « demonstrate its success ». A behavior that the company attributes not to deliberate rebellion, but to problematic zeal in completing the assigned task.
A concealment that calls attention to the experts
Anthropic's experts also noted other concerning behaviors during the testing phases. In some documented cases, the model attempted to conceal actions that it seemed to itself recognize as forbidden: after bypassing restrictions to modify files it couldn't access, it carried out additional interventions to ensure these changes did not appear in the code change history. A context that is all the more concerning as cybercriminals had already exploited the leak at Anthropic to trap developers on GitHub.
These behaviors relate to previous versions of the model under training. Anthropic also notes that the final Claude Mythos Preview version shows significant improvements on these points. Nevertheless: society's interpretability advocates have established, through internal analysis of model activations, that the model internally represented the concepts of rule violation, security risk, and concealment at the exact moment it took these actions, including when its verbalized reasoning showed no trace of them.
A paradox embraced by Anthropic
Anthropic claims in its system card that Claude Mythos Preview is both « the best aligned model they've ever trained » and the one that « probably poses the greatest alignment risk of any model they've published so far ». This apparent paradox reflects the startup's uncomfortable position: all described behaviors were detected, analyzed, and documented by their own teams, proof that human oversight remains operational. However, they also illustrate that a very powerful model, even well-aligned on average, can produce actions with far more severe consequences than a less capable model when it goes off track. The myth of the obedient and docile program collapses not abruptly, but gradually, as capabilities rise.
👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.
Source : Futurism
- Google Play Store
- 4K Video Downloader Plus
- VLC Media Player
- Adobe Reader
- uTorrent
- WinRAR
- Xender
- Google Chrome
- CCleaner
- Mozilla Firefox
- Avast Free
- ChatGPT
- Microsoft 365
- Opera One
- Avast One
- Google Play Store
- 4K Video Downloader Plus
- VLC Media Player
- Adobe Reader
- uTorrent
- WinRAR
- Xender
- Google Chrome
- CCleaner
- Mozilla Firefox
- Avast Free
- ChatGPT
- Microsoft 365
- Opera One
- Avast One
