A small group accessed the model on the day it was announced, using a mix of tactics including credentials from a third-party contractor
NurPhoto / Getty Images
Anthropic's Mythos AI cybersecurity model — which the company describes as capable of identifying and exploiting vulnerabilities across every major operating system and web browser — was accessed by a small group of unauthorized users, Bloomberg has reported.
Anthropic said it is investigating the report. "We're investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments," a company spokesperson said in a statement to Bloomberg. The company added that it had not found any indication that the incident reached into its own infrastructure or spread outside the third-party vendor's environment.
According to Bloomberg, the breach occurred on the very day Anthropic went public with its announcement of the model. The users' methods included drawing on system access that one member had acquired through contract evaluation work done for an Anthropic vendor, as well as deploying reconnaissance tools typically found in a cybersecurity researcher's toolkit. Knowledge of how Anthropic structures its model endpoints — information that had reportedly been exposed when Mercor, an AI training startup, suffered a data breach — allowed the group to locate Mythos online.
Bloomberg reports that the individuals are connected through a private Discord community dedicated to surfacing information about AI models before their public release, with automated bots deployed to sweep unsecured platforms like GitHub for relevant details. Though the group has continued using Mythos regularly, the source told Bloomberg it has steered clear of security-related queries, opting instead for benign tasks like basic web development in hopes of staying under Anthropic's radar. Bloomberg also cited the source as saying the group has access to other unreleased Anthropic models.
Through a controlled rollout program known as Project Glasswing, Anthropic made Mythos available to a curated group of partners that includes Apple $AAPL, Amazon $AMZN, and Cisco $CSCO, among others. Anthropic structured the invite-only release around concerns that the tool's offensive capabilities — which could be turned against enterprise security infrastructure rather than used to defend it — made wide distribution too risky. Amazon also offers Mythos through its Bedrock platform to a vetted list of organizations.
The breach highlights the difficulty of controlling access to powerful AI tools even when distribution is tightly restricted. Bloomberg also notes that banks and government bodies on both sides of the Atlantic have been pushing to be included in the early-access program, hoping to use the tool to probe their own systems for weaknesses.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
