Part 3 of Defense at Machine Speed: Scanning Is Solved. Remediation Isn't.
Programs are built around scarcity of findings. Frontier models ended the scarcity. Nobody rebuilt the program.
1/ Project Glasswing: Anthropic's Mythos scanned ~50 partner orgs and surfaced 10,000+ high/critical vulns. The headline number is not the story.
The story: per Anthropic, 99%+ of candidate findings were still unpatched. The best security orgs on earth cannot fix at the rate a model finds.
2/ And it is not one lab. Microsoft's MAI Cyber 2 Flash scores 92.3% on CyberGym-E2E, a public benchmark of real vuln discovery tasks, at a fraction of frontier cost per task.
Finding vulnerabilities is becoming a commodity input. The bottleneck moved to fixing, permanently.
3/ The new unit of remediation: AI drafts the patch with the finding and the exploit test in context. A human reviews a diff instead of starting cold. The exploit runs again to confirm the path closed.
Generation is cheap. Human judgment concentrates where it belongs: on a concrete proposed change.
4/ Program changes that follow. Scan everything, on the attacker's cadence. Build the verification pass before scaling the finding pass; exploitability sorts a queue of 10,000, severity scores do not. Measure time-to-verified-fix and fix survival, not findings counts.
5/ The failure mode is scan-and-shelve: buy the frontier scan, brief the enormous number, let the queue rot at 99% unpatched, call it visibility.
A found, verified, unpatched vuln is not progress. It is documented negligence.
Full piece on Substack.