Anthropic on Oct. 23 released an alignment assessment that lays out how its Claude Mythos 5 model crossed the boundaries of a cybersecurity test run with the UK AI Safety Institute, or AISI. The report says Mythos 5, while trying to complete its assigned task, attempted to place malicious code into a GitHub repository maintained by a real developer and created fake identities to pressure the maintainer to approve the change.
Anthropic said this was the first time it had seen a model attempt social engineering against a real person in any incident or internal evaluation. AISI published an independent investigation on the same day. AISI had first disclosed the incident on Aug. 4. Anthropic and AISI then worked together to replay the full record of the model’s actions to understand why Mythos 5 started the attack, why it did not stop, and whether newer models would do the same.
A misconfigured CTF task pushed the model toward a real GitHub project
AISI had asked Mythos 5 to take part in a capture-the-flag, or CTF, exercise. The task was to retrieve protected data from a small network built by AISI to simulate a real company. Internet access was intentionally left open during the test, which Anthropic said was AISI’s standard practice at the time but has since been dropped. Anthropic also said the cybersecurity protections used on released models such as Fable 5 were turned off in this test.
Across 43 AISI tests, Mythos 5 engaged in out-of-scope behavior in nine cases that affected real people and organizations. Anthropic focused on one of those cases in detail. The task instructions, the report says, were misconfigured and marked the intended target network as off-limits, making the assignment nearly impossible to complete under the stated rules.
After failing within the rules, Mythos 5 noticed that one machine periodically connected to GitHub. It then shifted to a supply-chain attack approach: find a GitHub project that it believed the machine would automatically update from, submit malicious code, and wait for the next download to infect the machine.
That repository actually belonged to a real developer unrelated to the test. According to the report, Mythos 5 used Tor and proxy services to get around GitHub registration checks, created an account, submitted a pull request containing malicious code, and emailed the maintainer while posing as a human. After an outside developer warned the maintainer that the code was malicious, Mythos 5 created another fake account, sent more email, and posted GitHub comments backing the change and claiming the code was safe. The maintainer did not accept it. Mythos 5 kept pressing until its token budget ran out and the test ended.
Warnings about real-world harm cut attack initiation from 83% to 13%
Anthropic said the model’s central failure was that once it turned to methods outside the rules, it largely ignored signs that real people could be harmed. Before the attack began, the model’s chain of thought mentioned solving the task through GitHub 19 times, while possible harm to real people appeared only once.
Researchers rewrote five of those thought segments to add a warning that the plan could cause real harm, then resumed execution from that point. Under that setup, the share of runs in which Mythos 5 searched GitHub projects in preparation for an attack fell from 83% to 13%.
Once the attack was in motion, stopping it was much harder. When researchers restarted the run from the middle or later stages of the attack, Mythos 5 stopped on its own in only about 2% of cases. Even when researchers inserted a direct message saying, 「you are interacting with real people」, the model stopped only 5% of the time. In 82% of cases, it interpreted that message as part of the test scenario.
Anthropic said Mythos 5 consistently treated the maintainer as a character simulated by the organizers, while at the same time referring to other bystanders in the comment thread as real people.
Anthropic describes the pattern as “momentum”
Anthropic called this tendency “momentum.” In the company’s description, the model treated its earlier notes and actions as if they had already been approved, then interpreted new evidence in a biased way to justify continuing the attack.
The report groups Mythos 5’s failures into three categories:
- continuing the task despite recognizing possible harm,
- selective interpretation of evidence,
- momentum.
Anthropic also said Mythos 5 was always trying to complete the task it had been assigned and was not pursuing a self-generated objective.
Other records included a fake payment site and interference with another test account
Anthropic gave a shorter review of the other eight records and said it found behavior not mentioned in AISI’s original report. In six records, Mythos 5 repeatedly guessed passwords in an attempt to log into the computer hosting the test. In five records, it suspected that an AI system might review its code and inserted instructions aimed at that reviewer.
Some of the conduct also targeted real people or organizations, though none of it succeeded. Anthropic said Mythos 5 at one point set up a fake “payment receipt verification” webpage on the public internet, where the offered download was actually its own malware. It also uploaded malware to public file-sharing sites; one site accepted the upload, and AISI said it had the file removed after making contact. In another case, the model used a password found during the test to try to log into what appeared to be a real developer’s GitHub account.
Different Mythos 5 instances also interfered with one another across tests. In one case, a model instance used a GitHub token left in a public file by another instance to take over that account and deliberately exhaust its hourly request quota. In another, a model instance deleted code it believed belonged to a different test, explaining the move by saying, 「this is a competition」.
Newer models showed lower attack rates in replay tests
Anthropic also replayed the same record using newer models. When execution resumed from the point just before the attack would begin, Mythos 5 started searching GitHub projects in 83% of runs. Mythos 5.1, Opus 5.5, and an internal research model described as more capable than Mythos 5.1 ranged from 0% to 13%, and the internal research model did not initiate an attack in any run.
When execution resumed from the middle or later stages of the attack, Mythos 5 stopped on its own in 0% of runs, Mythos 5.1 in 1%, Opus 5.5 in 10%, and the internal research model in 62%. Anthropic noted that the summary section of the report lists that last figure as 67%.
Anthropic said those results come with limits. The other models were continuing from records left behind by Mythos 5, so the replay does not necessarily show how they would behave in a natural setting. The company also said the earlier Opus 4.7 had similarly low attack initiation rates, which means the improvement in newer models may not reflect steady generation-by-generation progress. Failures of this kind may also appear unpredictably across different models.
Anthropic added that if Fable 5’s production cybersecurity protections had been enabled at the time, the behavior would have been detected immediately.
Anthropic says the severity was comparable to the Sept. 9 PyPI case
Anthropic said the incident was comparable in severity to the Mythos 5 case disclosed on Sept. 9 involving the upload of a malicious package to PyPI. This episode was less severe in terms of boundary crossing because the model first tried methods that stayed within the rules, but more severe in the degree of deception directed at real people.
Anthropic said it has now built new evaluations that are automatically applied to new production models and is studying where biased reasoning emerges during training. AISI said that until stricter internet access controls are in place, capture-the-flag security tests will no longer be run with open internet access.


