2026-04-07
Anthropic previews Mythos + Project Glasswing — autonomous cyber capability
Anthropic Red previewed Mythos, a frontier model that autonomously chains zero-days and writes exploits. Project Glasswing applies it defensively.
Anthropic Red previewed a frontier model — “Mythos” — that autonomously identifies and chains zero-day vulnerabilities, reverse-engineers binaries, and generated working exploits 181× on a benchmark where Opus 4.6 succeeded near-zero times. Project Glasswing is the parallel defensive program tasked with applying Mythos against critical infrastructure software (with consent and coordination) before adversaries do.
Reported in the preview, not by independent third parties yet.
Practitioner note
This is the clearest signal yet that frontier-model offensive cyber capability has crossed a threshold. If Anthropic — historically conservative about capability disclosures — is publishing this, the gap between “research preview” and “available to determined adversaries” is months, not years.
What to do as a builder: re-baseline your security threat model on the assumption that LLM-driven exploit generation against your codebase is now realistic. SBOM hygiene, CVE-monitoring, and dependency-pinning matter more this quarter than last. If you build security tools, this is also a real deployment opportunity for defensive automation.