Tom's Hardware
US hardware and technology journalism site; challenged Anthropic's Mythos zero-day claims, finding they rested on only 198 manual reviews.
On 9 April, Tom's Hardware challenged Anthropic's claim that Mythos had found thousands of zero-days, reporting that it rested on 198 manual reviews.
Last refreshed: 24 September 2026 · Appears in 2 active topics
Did Tom's Hardware's Mythos critique hold up when AISI published its independent evaluation?
Timeline for Tom's Hardware
Mentioned in: Stargate and Atlas top Argentina data-centre pipeline
Data Centres: Boom and BacklashMentioned in: OpenAI cuts compute target by $800bn
Data Centres: Boom and BacklashMentioned in: AISI confirms Mythos 20-hour attack chain
AI: Jobs, Power & MoneyMentioned in: Tom's Hardware challenges Mythos zero-day claims
AI: Jobs, Power & MoneyMentioned in: Anthropic withholds Mythos from public release
AI: Jobs, Power & MoneyBackground
Tom's Hardware is a US technology journalism publication focused on hardware reviews, benchmarks and technical analysis. Its coverage includes processors, graphics cards, memory, storage and related technology for enthusiast and professional IT readers.
The publication's relevance in security reporting comes from applying technical scrutiny to claims that affect hardware, software and computing practice. On 9 April, it examined the evidential basis for Anthropic's Mythos zero-day claims rather than treating the marketing figure as settled.
That role is distinct from that of a regulator or model developer. Tom's Hardware reports and assesses technical evidence; it does not set safety policy or Conduct official evaluations. The AISI evaluation published on 15 April supplied an independent assessment of Mythos's attack-chaining capability.
Tom's Hardware tested Mythos claims
On 9 April, Tom's Hardware reported that Anthropic's claim of thousands of zero-days found by Mythos rested on 198 manual reviews. It also noted that many flagged flaws were in software no longer in active use, narrowing what the headline figure established.
The UK AI Security Institute's 15 April evaluation gave the publication's scrutiny a sharper frame. The Institute did not find single-task superiority over public frontier models, but reported that Mythos completed a 32-step attack chain estimated at 20 hours of trained-human work. For Tom's Hardware, the distinction matters: a disputed zero-day count did not settle the separate question of autonomous attack capability.