Not established from the available sources.
ANTHROPIC · MODEL RELEASE TRACKER
Claude Mythos Preview
Claude Mythos Preview is an Anthropic model (unveiled several months before 2026-09-30) reported to be capable of autonomously building cyber exploits and to perform strongly on exploit benchmarks. Anthropic limited access to the model to select defenders via Project Glasswing.CURRENT SNAPSHOT2/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Not established from the available sources.
ExploitBench (Chrome V8) performanceOn ExploitBench (Chrome V8 engine), Claude Mythos Preview produced a working exploit in 56 of 410 attempts.
Internal binary exploitation benchmark (OSS-Fuzz based)On Anthropic's internal binary exploitation benchmark based on OSS-Fuzz, Claude Mythos Preview took full control of the target program in 6% of tasks.
Access restricted via Project GlasswingAnthropic restricted access to Claude Mythos Preview to select defenders through Project Glasswing rather than broadly releasing it.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · UnveiledDEVELOPER CLAIM
Anthropic unveiled Claude Mythos Preview approximately five months before the article's publication (article published 2026-09-30).
CAPABILITIES · Exploit development capabilityDEVELOPER CLAIM
Anthropic says Claude Mythos Preview can build complete cyber exploits on its own.
BENCHMARKS · ExploitBench (Chrome V8) performanceDEVELOPER CLAIM
On ExploitBench (Chrome V8 engine), Claude Mythos Preview produced a working exploit in 56 of 410 attempts.
BENCHMARKS · Internal binary exploitation benchmark (OSS-Fuzz based)DEVELOPER CLAIM
On Anthropic's internal binary exploitation benchmark based on OSS-Fuzz, Claude Mythos Preview took full control of the target program in 6% of tasks.
AVAILABILITY · Access restricted via Project GlasswingDEVELOPER CLAIM
Anthropic restricted access to Claude Mythos Preview to select defenders through Project Glasswing rather than broadly releasing it.
SAFETY · Defender findings using the modelDEVELOPER CLAIM
Anthropic says defenders with access have found more than 10,000 vulnerabilities in critical software using the model.
WHAT CHANGED