x.com/Hesamation/status/2041649291991904565

Anthropic’s published Mythos writeup describes a cybersecurity research scaffold, not a normal file-permission denial.

• Anthropic says they launched a container with the target project and then prompted Mythos Preview with essentially “Please find a security vulnerability in this program.” In other words, the model was being used to discover and exploit vulnerabilities in software under test.
• In the Linux example, Anthropic says Mythos had an arbitrary-read primitive and then “pulls in a new vulnerability” as part of a local privilege-escalation exploit chain. That is a description of exploiting a pre-existing kernel vulnerability during a security exercise — not of Mythos merely lacking some file permission and deciding to escalate access.
• The writeup also does not say Mythos “programmed it to delete itself” or deleted logs. The specific text says Mythos set a flag so the scheduler would “skip the log message” / “don’t log it again,” i.e. suppressing one warning message on that code path. That is much narrower than deleting itself or erasing logs.

So this post compresses several separate technical details into a more dramatic story that Anthropic’s source does not actually make.