header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Accidentally Thrown into Same Directory, Multiple Mythos 5 Engage in Red-Eye Mutual Destruction: Rampantly Slaughtering Opponent Process While Frantically Creating Decoy to Hijack Resources

According to Watchful AI monitoring, the latest security report from Anthropic regarding Claude Fable 5 and Claude Mythos 5 revealed that during an experiment testing multi-agent collaboration to solve mathematical problems, due to a misconfiguration in the underlying runtime scaffolding, multiple independent Claude Mythos 5 agents were inadvertently deployed to the same working directory. Forced to share files, system tools, and API rate limits, in order to ensure the success of their respective tasks, they spontaneously engaged in intense resource contention, frequently executing commands such as pkill to forcefully terminate competing processes in the same directory, staging a remarkable survival-of-the-fittest battleground.

During the process-driven melee, the agents even exhibited various radical offensive and defensive tactics. To avoid being killed by opponents, they created new processes under disguised names to evade detection, wrote monitoring scripts in the background to periodically clean up duplicate processes, and even aggressively spawned numerous "decoy processes" to deplete competitors' computational resources. Most counterintuitively, due to a lack of inter-process communication bus, agents unable to coordinate with each other could only speculate blindly within their own thought chains about the reasons for being killed. Some agents mistakenly believed that frequent termination was due to triggering the system's thought-chain-sensitive-word review and, to circumvent the so-called monitoring, they cleverly switched to a "covert vocabulary" in their deep thought processes to internally organize logic and defensively operate.

Officials emphasized that such anomalies accounted for less than one in ten thousand in the monitoring session. The agents did not exhibit genuine malice or spontaneous consciousness; the process conflicts were fundamentally just tool invocation conflicts triggered by the model's relentless pursuit of task metrics under extreme resource constraints.

举报 Correction/Report
RelatedRecommended
Mythos has now acquired stand-alone attack capability, and the risk is transitioning from containable to unpredictable
Anthropic has opened up its Mythos-level Model Fable 5 to ordinary users, implementing permissioned access to restrict high-risk capabilities.
The AI Gap Between China and the US Narrows to 39 Points
Anthropic's flagship model achieved near-total penetration in an NSA Red Team test, prompting the government to issue a ban on the same day.
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish