Mozilla used Anthropic’s Mythos Preview to find and fix 271 vulnerabilities in Firefox — a stark example of the model’s ability to spot software flaws. Still, a small group of Discord users managed to reach Mythos despite Anthropic’s tight rollout controls. They traced a path to the model using data from a Mercor breach plus existing contractor permissions, and they reportedly limited their activity to building simple websites. The episode shows how leaked configuration data and predictable hosting patterns can defeat gated access to powerful AI tools.
How the access happened
Anthropic kept Mythos Preview behind restricted gates because the company says the model can find security vulnerabilities in code and networks. But the restriction didn’t stop a handful of amateur investigators on Discord from getting in. They started with material leaked in a breach at Mercor, an AI training startup that works with developers. Using that data, the group made an educated guess about where Anthropic hosted the model online based on how Anthropic has formatted access to other models.
One member of the Discord group also had permissions from work with a contracting firm that provides services to Anthropic. That access let the user reach other Anthropic models; combined with the leak-driven sleuthing, the group gained entry to multiple unreleased Anthropic models, according to reporting on the incident.
What they did with Mythos
The people who reached Mythos didn’t appear to try to weaponize the model. Instead, they used it to create simple websites. That choice was reportedly deliberate: simple web projects are less likely to draw attention than obvious attempts to probe security tools or scan networks for vulnerabilities. The Discord users appear to have focused on staying under the radar rather than stress-testing Mythos’s fault-finding power.
Anthropic has said it’s being careful about who it lets use Mythos because the model can surface vulnerabilities. Mozilla’s use of Mythos Preview to hunt for problems in Firefox and the subsequent patching of 271 vulnerabilities is the clearest public example of the model’s potency. At the same time, the Discord incident shows that human investigation combined with leaked metadata can bypass gatekeeping built around capability-based risk.
Where controls failed
The chain that led to unauthorized access had at least two weak links. First, Mercor’s leaked data provided clues about how Anthropic organizes and exposes its models. The Discord group didn’t break cryptography or run elaborate exploits; they followed a trail of information and inferred an online location.
Second, contractor-level permissions already held by one user gave the group a bridge from public sleuthing to internal resources.
That pattern — exposed configuration details plus legitimate credentials — is a common breach vector. In this case, it meant a tight release plan for an advanced AI model didn’t stop unauthorized use. Anthropic’s safeguards were effective against blunt-force attacks but less so against low-effort detective work combined with legitimate access rights.
Security research and real-world stakes
Mythos is framed inside the industry as a double-edged sword. Its ability to find vulnerabilities makes it useful for defensive security work and patching, as Mozilla’s results show. But the same ability also raises concerns about misuse: a model that can find flaws could be used to locate and exploit zero-day vulnerabilities if it fell into the wrong hands.
That risk is why Anthropic limited access. It's also why the Discord group’s choice to build benign websites matters: they avoided behavior that would make detection more likely. Their caution reduced the chance Anthropic’s monitoring systems would flag unusual requests tied to vulnerability scanning. The episode exposes a tension in AI governance — tools that can help patch vulnerabilities can also be misused if access controls fail.
Related Articles
- Google unveils AI cyber agents after $32B Wiz acquisition
- Meta to Run AI Workloads on Millions of AWS Graviton CPUs
- Rocket Report: New Glenn setback, Artemis III readies
Mozilla said it used Mythos Preview to find and fix 271 vulnerabilities in Firefox 150.
This article was created with AI assistance.
