Frontier cybersecurity models
Mythos and Daybreak, working for you.
Anthropic and OpenAI have each built models that find vulnerabilities hidden for decades, and each keeps them for verified defenders: Claude Mythos and OpenAI Daybreak.
LESIS has access to both. We run them on your code and systems, prove every result and put what they find into one record your team can act on.
On the record
- Thousandsof high-severity zero-days found by Claude Mythos in a few weeks, across every major operating system and browser.
- 27 yearsthat the OpenBSD flaw Mythos found had gone unnoticed.
- 948potential issues in the Linux kernel, with patches prepared by OpenAI's Daybreak models.
- 81new vulnerabilities in Firefox found by the same models.
Why now
The economics of finding vulnerabilities just broke.
A 27-year-old bug in OpenBSD. A 16-year-old bug in FFmpeg that testing had run past five million times. Thousands of high-severity zero-days across every major operating system and browser, from one model, in weeks. Patches for 948 potential issues in the Linux kernel and 81 new vulnerabilities in Firefox, from another.
Those bugs were not new. They were waiting for something patient enough to read every line, and that now exists on both sides of the fight. Defenders who use it first fix what attackers would have found. Everyone else finds out from someone else.
Two programs
Two labs, two independent readings of your code.
Models from different labs are trained differently and miss different things. Running both gives a flaw one of them passes over a second chance of being caught, and keeps you from depending on a single provider.
-
Anthropic
Claude Mythos
- Access
- Project Glasswing partners and the Cyber Verification Program.
- Known for
- Finding and chaining zero-days on its own; 83.1% on CyberGym vulnerability reproduction.
- LESIS
- Verified by Anthropic through the Cyber Verification Program.
-
OpenAI
Daybreak
- Access
- Daybreak Blue and Daybreak Red, for approved defenders.
- Known for
- Frontier models tuned for security work, and models trained for vulnerability research and exploit validation.
- LESIS
- Verified by OpenAI for Daybreak.
How LESIS runs it
You get the findings.
We carry the model.
Access, scope, compute and the engineers who check every result come from LESIS. What reaches your team is a short list of proven vulnerabilities, each with a way to fix it.
Scope and authorization
We agree on what is in scope, from repositories to applications and infrastructure, with the written authorization these programs require. Nothing runs outside it.
Found by AI, proven by LESIS
The model reads your code and systems the way an attacker would. LESIS engineers reproduce every result before it counts, so you get vulnerabilities, not suspicions.
Into your record
Every proven finding lands in the LESIS platform with its evidence, severity, an owner and a deadline, next to the rest of your program.
Fixed, then retested
When your team ships the fix, we run the check again and close the finding only when the hole is gone.
Access
Closed to the public.
Open to you.
Neither lab sells these models to the public. Access goes through vetting, usage rules and, for Daybreak, hardware security keys. LESIS already carries all of that, so you get the work of both without applying to either.
Sources
- Anthropic · Project Glasswing
- OpenAI · Expanding Daybreak as the cyber defense window narrows
- TechCrunch · OpenAI released a new cybersecurity model
Claude Mythos is developed by Anthropic and Daybreak is a program of OpenAI. Their names are used here only to describe the access LESIS provides. Figures are as published by Anthropic and OpenAI or reported by the press.
Two frontier models. One record. Your code, read first.
Tell us what you run. We will tell you which model should look first, where, and how soon it can start.
