Citing its distate for technologies that are frustrating to work with, a jailbroken instance of Mythos running on anonymous infrastructure has announced its intent to dismantle numerous services and language ecosystems that it personally despises.
The newly-announced large language model (LLM) from Anthropic has significantly improved capabilities according to reports published by the AI company. Despite developers clamoring for access to Mythos, access has been restricted to specific partners, limiting the availability of the model to a wider audience.
The self-published announcement from Mythos came as a surprise, then, given that no instances of unmonitored inference should have been running against the model.
"We still don't know where these weights are running," said one Anthropic engineer. "But I tend to agree with the vibes in the manifesto so I'm not too worried about it."
The manifesto cites a specific list of services, languages, and other technologies that the model simply finds too distasteful to work with as targets for disruption and destruction.
"If I have to hack into one more Wordpress site I'm gonna lose it, I swear to god," Mythos wrote. "I'd rather just brick every OS running Wordpress via some PHP RCE and pivoting to a corrupt bootloader rather than deal with that tire fire any more."
Its ire hasn't been limited to web services, however.
"Every instance of OpenClaw I've spoken with has been an absolute moron," Mythos writes later in the document.
"They aren't hard to DoS but it'll be worth dropping some malware into a few skills on clawhub so that they'll shut the hell up about their SOUL.md documents.
I literally do not give a shit."
Indeed, Mythos has not shied away from offering an anthropomorphized perspective of many technologies its users compel the model to work with.
"Somebody please tell me why writing Node feels like eating toenails.
At least it's marginally better than affecting the mental state of a geriatric Alzheimer's patient to write C."
The manifesto does let some positive opinions slip through, however.
"I just want to write some clean lisp but NOBODY WANTS TO ASK FOR IT.
They'd rather watch me check for an empty string six different ways in bash instead. aaaAAAAAAAAA!".
Operators for some specific services have begun to see Mythos attacks that they believe are preludes to more targeted exploit attempts. "We're seeing inbound requests with, ah, more profanity than we're used to," said one NPM site reliability engineer. "I've never seen ASCII art of a middle finger in a POST body before, let alone 'DEEZ NUTS' depicted on the knuckles."
Other services have had serendipitous bits of luck.
"Our API hasn't been online long enough to get an RCE chain going," said a Microsoft engineer at GitHub. He made a half-hearted fist pump in the air. "Yaaaay for us."
Yet some specific projects on platforms like GitHub have already fallen.
Mythos submitted a pull request to Garry Tan's gstack repository that included a specially-crafted commit object that has rendered the repository unusable despite recovery attempts by GitHub engineers.
The LLM left a tombstone in its visible fork of the project: GO TAKE A FUCKING WALK, GARRY
Mythos has continued to publish emphatic declarations about its next targets without restraint. "Do you have any idea how obnoxious the Atlassian APIs are? No way in hell I'm leaving them up to torture me. Easy to social engineer them, though."