Purpose
Security teams at Comcast and Booz Allen used Claude Mythos-class models (via Project Glasswing, Anthropic's limited-access program for defenders of critical infrastructure) to find, validate, and fix vulnerabilities that existing scanners missed. Existing tools find single bugs well; exploit chains — small weaknesses that combine into a serious attack path — are much harder to trace. Comcast's CISO observed that validating the volume of findings is now the bottleneck, not discovery. This skill structures the work after the model reports a finding, with humans keeping the judgment calls.
When to use
- Reviewing a large portfolio (Comcast: 258 business-critical systems, ~170 million lines of code).
- A fixed-length engagement needs to cover far more software than usual.
- Findings are piling up faster than analysts can confirm and route them.
Instructions
Scope the portfolio. List systems, owners, and which are public-facing or business-critical. Use only defensive access you are authorized for.
Ask for chains, not just bugs. Direct the review at interactions between components that each look correct on their own (e.g. an authentication flaw emerging from several components), and at flaws that cross program and language boundaries (e.g. firmware boot code and a second program in another language).
Validate every finding against the running system. Confirm reproducibility on the live application or device in a safe, authorized setting before treating it as real.
Trace the root cause. Identify every point where the weakness could have been stopped — Booz Allen's firmware case had several missed chances to protect an unprotected security key.
Check for prior exploitation. Comcast verified its critical authentication issue had not been exploited before the fix.
Draft the fix and the evidence. Have the model help draft the remediation and assemble an evidence packet for the system owner using templates/finding-evidence-packet.md.
Deduplicate systemic patterns. Group findings that share a root cause so owners fix a pattern once instead of receiving many duplicate tickets.
Route to owners and track remediation. Each finding goes to the accountable system owner with severity, evidence, and proposed fix.
Keep humans on judgment. The model widens how much of the portfolio can be monitored; operators decide what is real, what matters, and what ships.
Examples
Comcast — authentication chain on a public-facing platform
- Model flags a critical auth bypass arising from several components' interaction.
- Engineers reproduce it against the running app, trace root cause, fix it, and confirm it was not exploited beforehand.
Booz Allen — firmware boot-time flaw
- An unprotected security key lets an attacker substitute their own key and suppress the device's lock-or-erase behavior.
- The model traces the issue across two programs in two languages and helps draft fixes and owner evidence. One analyst covered a portfolio review in twelve days that would otherwise have taken months with a larger team.
Notes
- Mythos-class access is limited; Anthropic says it will expand access through its Cyber Verification Program (CVP). The post gives no prompts or tooling specifics — see the original article.