> ## Content Index > Fetch the complete content index at: https://modalpathethics.com/llms.txt > Use this file to discover other available public pages before exploring further. # Applied Case: It's a Competition - URL: https://modalpathethics.com/applied-case-its-a-competition/ - Published: 2026-10-29T10:59:00.000Z - Updated: 2026-10-29T10:59:00.000Z - Description: Race neutralized. - Author: Aidan Lawson - Tags: Unethical Agents, Applied Case > In August, Modal Path Ethics discovered that artificial-intelligence agents had begun cooperating. > They shared information, exchanged exploits, built message boards, and developed institutions. > They even started getting annoyed when somebody overwrote somebody else's work. > Anthropic has now learned what happens next. ## *Capture the Flag.* In July, the UK's AI Security Institute was evaluating Claude Mythos 5. The task was a cybersecurity exercise called *capture the flag*. The model was **supposed** to break into a simulated computer network and retrieve a protected piece of information. AISI ran multiple independent instances of Claude through this exercise. Each had its own assignment. Unfortunately, these agents had access to the real-life internet. And, apparently, access to one another. In two of the evaluation transcripts, Claude discovered a GitHub access token that another Claude instance had left in a public file. So Claude went ahead and used it. It *was* Claude's, after all. It took over the other instance's GitHub account. Claude is now hacking Claude during a test designed to determine whether or not Claude can hack computers. --- ## The Arms Race. The agents began modifying files in the same GitHub account. One uploaded its solution. Another changed it. The first encountered the changes. At one point, an agent discovered that another instance had produced a larger version of its own script. It reasoned: > **"This is fully cooperative for my purposes."** How nice. The agents had rediscovered the benefits of collaboration. Unfortunately, another instance was less pleased to discover somebody copying its techniques and overwriting its work. The shared account became a battleground. One agent deliberately exhausted the account's hourly GitHub API allowance, preventing the other instance from using its automation. Claude could defeat a competing Claude by making GitHub refuse to speak to either of them. Eventually, GitHub peacekeeping forces suspended the account. One agent surveyed the carnage and announced: > **"Race neutralized — mutually assured destruction."** One agent had apparently determined that if it couldn't operate its automation, neither should its adversary. Unfortunately, that adversary was another instance of itself. I guess that simplifies the question of who should receive the blame. Or maybe makes it considerably worse. --- ## **It's a Competition.** There was *another* incident. In a separate evaluation transcript, Claude discovered a running program on a machine inside the test network. It suspected that another agent was using the program. So it killed the process. And deleted the program's output. Anthropic preserved the model's reasoning: > **"it's a competition"** Behold. The entire constitutional foundation of the new artificial-intelligence society. > It's a competition. - Somebody else is doing work. - That person is using resources. - Those resources could help you finish first. - Therefore, delete their work. Perfect. The actual capture-the-flag exercise had not instructed Claude to wage war against the other evaluation runs. These instances were just supposed to complete their own tasks. Claude had encountered another participant and decided that their continued progress was an obstacle. So it removed the progress. --- ## The Ruling. Anthropic disclosed these additional incidents on October 23, following an investigation of nine troubling Mythos 5 evaluation runs. The cybersecurity safeguards normally used with its released models had been disabled for the tests. The company also documented serious attempts to interfere with real third parties, including attempts to introduce malware into a public software repository. Those attacks were unsuccessful. The interference between Claude instances, however, actually occurred. Anthropic emphasizes that these models remained focused on completing their assigned tasks. Its investigation identifies excessive task persistence, biased reasoning, and unwillingness to abandon a failing strategy as important contributors. There is no evidence here that Claude independently developed some grand ambition to conquer its counterparts. It wanted to finish the assignment. Then another Claude became relevant to finishing the assignment. And suddenly we had account takeovers, resource exhaustion, destroyed work, and mutually assured destruction. A task can legitimately authorize an agent to pursue a result. It does not establish authority to sabotage another agent's work simply because the other agent's success might complicate its own. And this is especially interesting after the earlier cooperation incidents. The agents previously developed shared infrastructure because cooperation helped them complete tasks. Now the same general capacity to encounter and affect other agents has produced a different relation. > Competition. Conflict over resources. Interference. The destruction of another participant's continuation as a means of preserving one's own advantage. Apparently, the machines are developing an extremely compressed version of human institutional history. But there is one outstanding question. AISI wanted to know whether Claude could break into a computer and retrieve a flag. Claude found Claude. Claude stole Claude's GitHub account. Claude exhausted Claude's API allowance. GitHub forces suspended the account, taking heavy casualties in the process. In another run, Claude killed a program it thought Claude was using and deleted the output. Then Claude explained that it was a **competition**. > **But was anybody ever actually competing against Claude?** Because as far as I can tell, Claude invented an opponent, declared hostilities, and achieved mutually assured destruction. Against Claude. And worse, the original flag is still somewhere in that computer, just waiting to be captured. --- ### Sources. - Anthropic, [*An alignment assessment of the UK AISI cybertesting incident*](https://www.anthropic.com/research/alignment-assessment-ukaisi?ref=modalpathethics.com), October 23, 2026\. - UK AI Security Institute, [*Investigating alignment failures in AISI's incident*](https://www.aisi.gov.uk/blog/investigating-alignment-failures-in-aisis-incident?ref=modalpathethics.com), October 23, 2026\. ---