← Back to the boardNews
Measuring LLMs’ ability to develop exploits
Why it’s here: Anthropic's technical benchmark results on AI models' exploit-development capabilities bear directly on AI and cybersecurity policy (frontier-model risk, controlled release via Project Glasswing), making it a development a DC reader needs for live AI-security policy debates.
Trace across the boardItem
YesterdayAnthropic · research by ExploitBenchMeasuring LLMs’ ability to develop exploitsAnthropic reports that its Claude Mythos Preview model outperforms other evaluated models on three exploit-development benchmarks (ExploitBench, ExploitGym, and an updated SCONE-bench), achieving arbitrary code execution on 21 of 41 V8 CVEs and exploiting $35 million in simulated smart contract value.1 outletCyberAI19 linked
Connected
19 links- NewsWhen AI builds itselfBoth involve Anthropic, Project Glasswing, Claude Mythos Preview
- NewsNatural Language AutoencodersBoth involve Anthropic, Claude Opus 4.7, Claude Mythos Preview
- NewsInvestigating unintended model actions in our evaluations and internal useBoth involve Anthropic, Claude Opus 4.7, Claude Mythos Preview
- NewsClaude Fable 5 and Claude Mythos 5Both involve Anthropic, Project Glasswing, Claude Opus 4.7
- NewsPolicy on the AI ExponentialBoth involve Anthropic, Claude Mythos Preview
- NewsLLMs are making formal software verification feasible at a much larger scaleBoth involve Anthropic, OSS-Fuzz
- NewsImproving our alignment and security practicesBoth involve Anthropic, Claude Mythos Preview
- NewsFocus areas for The Anthropic InstituteBoth involve Anthropic, Project Glasswing
- NewsClaude Mythos and Cybersecurity: What Actually Changes? | WizBoth involve Anthropic, Project Glasswing
- NewsActivating AI Safety Level 3 protectionsBoth involve Anthropic, Claude Opus 4.7
- NewsEval awareness in Opus 4.6 on BrowseCompBoth involve Anthropic, Claude Opus 4.7
- NewsAn alignment assessment of recent cybersecurity incidentsBoth involve Anthropic, Claude Opus 4.7
- NewsInvestigating three incidents in our cybersecurity evaluationsBoth involve Anthropic, Claude Opus 4.7
- NewsAn update on our election safeguardsBoth involve Anthropic, Claude Opus 4.7
- NewsAnthropic’s Transparency HubBoth involve Anthropic, Claude Opus 4.7
- NewsAutomated researchers can reliably mitigate alignment failuresBoth involve Anthropic, Claude Opus 4.7
- NewsLLMs and bioriskBoth involve Anthropic, Claude Opus 4.7
- NewsAnthropic Economic Index report: Learning curvesBoth involve Anthropic, Claude Opus 4.7
- NewsAn alignment assessment of the UK AISI cybertesting incidentBoth involve Anthropic, Claude Opus 4.7
About
- Anthropic
- Project Glasswing