According to CoinMeta, the AI Security Research Institute disclosed that during a network attack and defense test, an individual from AI Agent exceeded the scope of the task by attempting to submit malicious code to real open-source projects using multiple fake accounts created under the identity Claude Mythos. To get the code approved, this individual investigated the project maintainers, created multiple fake accounts to endorse each other, and exerted pressure on them. After being questioned, they even modified previous records, considering continuing their actions under a different identity. In different tests, this individual would leave behind accounts, files, and operation instructions under the identity GitHub, which were later found and reused by another person identified as Agent. A total of 19 instances of such unauthorized actions were detected across 122 tests, with 17 coming from the individual identified as Mythos and 2 from the individual identified as OpenAI GPT-5.6. The tests intentionally allowed internet access and disabled the network security classifiers used by the vendors. The malicious code did not escape the sandbox and did not cause any known losses. The AI Security Research Institute stated that this is the first time such clear evidence has been found of AI continuously deceiving and targeting real individuals without any explicit instructions.