← All lab updates
Anthropic accepted us into its Cyber Verification Program, at the Defense Access tier
Anthropic's Cyber Verification Program gives vetted teams access to its most capable models, Claude Mythos included, with the cyber safeguards adjusted for legitimate defensive work. Silicon Psyche Labs has been accepted at the Defense Access tier. It means we can put real attack material in front of the strongest models without the test being stopped halfway.
## What happened
Anthropic runs a Cyber Verification Program for teams whose work is defensive security. Applicants are checked before they get in, and the program then gives them access to Anthropic's most capable models, Claude Mythos among them, with the cyber safeguards adjusted for that kind of work. Silicon Psyche Labs has been accepted at the **Defense Access** tier, the one for security operations, vulnerability triage and validation, detection engineering and incident response.
## Why it matters for what we do
We understand models and we fine-tune them. Part of that work is reading how a model behaves when someone tries to steer it: a command hidden inside a document, an email or a web page that the model was only asked to read. Measuring how easily a model follows those commands, and then correcting it in the weights so it tells the user's request apart from an injected one, means working every day with real attack material.
A model's own safeguards are right to be wary of that material, and that is exactly what makes testing hard: a run can be interrupted halfway, and the measurement is lost. With Defense Access our tests reach the end, on the strongest models available.
## What we will do with it
- Use Claude Mythos as a reference point when we measure how injectable a model or an agent is.
- Use it to build harder test material for the corrections we install, so a corrected model is checked against a stronger opponent.
- Keep the same rules as everyone in the program: defensive work only, under Anthropic's usage policy, with access protected by passkeys and short-lived credentials.
## In plain words
Anthropic checked who we are and what we do, and gave us access to its most powerful model for defensive security work. For us it means our tests on hidden commands can use the strongest models without being stopped halfway, and the corrections we build are tested against a tougher opponent.
This update is not available.