Anthropic’s S-1/A says Claude’s score on ExploitBench, “a benchmark developed by Carnegie Mellon researchers that evaluates AI models’ ability to exploit software vulnerabilities”, went from 40% with Opus 4.8 to 78% with Mythos 5 and 91% with Opus 5.5, “over a period of approximately four months”.
That is a gain of 51 percentage points, about 2.3 times the starting score, by our arithmetic. A slide from Anthropic’s IPO roadshow video, posted on X by Tanay Jaipuria on 11 November, marks cyber’s threshold as crossed in April 2026, five months after coding’s.
Business Pill · Least Privilege
An evergreen explainer on giving an AI agent only the access its task needs. It teaches the general idea only and does not discuss Anthropic’s filing.
The key insight: Anthropic’s S-1/A says Claude’s ExploitBench score rose from 40% with Opus 4.8 to 78% with Mythos 5 and 91% with Opus 5.5 in about four months, a 51-point gain by our arithmetic. The roadshow video places cyber second after coding, and a slide marks its threshold as crossed in April 2026.
The filing puts the scores in an argument about adoption: “When a model crosses the capability, reliability, and trust threshold required for a category of work, adoption in that category generally moves quickly from experimentation into production”.
It names coding as the first such category and adds: “Cybersecurity is following a similar curve”. After the scores it says: “Cyberdefenders are now identifying and patching vulnerabilities at scale”.
By our arithmetic, the step from Opus 4.8 to Mythos 5 added 38 points and the step to Opus 5.5 another 13. The filing gives the four-month span for the whole run, not dates for each test.
Claude’s ExploitBench scores by model over about four months, as reported in Anthropic’s S-1/A of 10 November 2026. ExploitBench is a Carnegie Mellon benchmark of AI models’ ability to exploit software vulnerabilities.
What Was Posted on X
Some slides from the Anthropic roadshow video
– big emphasis on fungible compute, ability to use compute efficiently and onboard new chips quickly – still very early in agents outside of coding – long-term target of 35% adjusted operating margins
Tanay Jaipuria posted on 11 November at 01:37 UTC four images the post calls slides from Anthropic’s roadshow video, including one titled “Making fast progress across knowledge work” that marks the cyber threshold as crossed in April 2026, after coding in November 2025. This piece sets that slide against the ExploitBench scores in Anthropic’s S-1/A.
What the Roadshow Video Adds
The slide, titled “Making fast progress across knowledge work”, shows coding at “Nov ’25 Threshold crossed” and cyber at “Apr ’26 Threshold crossed”. Its footnote defines the threshold as “the release after which long-running, multi-agent autonomous work became the frontier default” and calls the chart an “Illustrative curve”.
The video’s official captions, which do not name speakers, say: “After coding came cyber, and with Mythos and Project Glasswing, Claude has helped cyber defenders identify and patch vulnerabilities at a societal scale”. They cite Mozilla: “In the past, they used to fix over 20 security bugs per release, and with Mythos Preview, they identified and fixed 271”.
The S-1/A says Mythos “runs with safeguards calibrated for vetted organizations and individuals doing cybersecurity and life sciences work” and is offered through trusted-access programs, including its Cyber Verification Program. It says the class first launched “in April 2026 as Claude Mythos Preview through Project Glasswing”.
Opus 5.5, the 91% model, is described as “the enterprise daily driver”. The filing also discloses misuse: in September 2025, it says, a Chinese state-sponsored group “manipulated our Claude Code tool into attempting infiltration into roughly 30 global targets”.
As we read the filing, the scores are offered as evidence for a pattern, that adoption moves fast once a model crosses a threshold for a category of work, with coding first and cybersecurity next.
As we read the access terms, the filing pairs its exploit scores with tiered access: the underlying model ships as Mythos for vetted cyber and life-sciences users and as Fable with production safeguards, while Opus 5.5, the 91% model, is a separate class the filing calls the enterprise daily driver.
Anthropic, S-1/A Amendment No. 2, 10 November 2026
“Cybersecurity is following a similar curve”
Three Implications
FOR THE SCORES 40%, 78% and 91% on ExploitBench over about four months, the S-1/A says; 51 points and about 2.3 times the start, by our arithmetic.
FOR DEFENDERS Mozilla used to fix over 20 security bugs per release and fixed 271 with Mythos Preview, the roadshow video’s captions say.
FOR ACCESS Mythos runs with safeguards for vetted organizations through trusted-access programs, including the Cyber Verification Program, the filing says.
The analysis says of Anthropic: “It is becoming the first major AI lab actively attempting to define the governance layer of the AI stack.”
As we read it, the filing pairs its highest exploit score with access rules, vetted users for Mythos and safeguards for the general release, which is the layer the analysis describes. The analysis does not discuss the filing.
What Is Not Established
The S-1/A text we searched does not give ExploitBench’s version, task count or test settings, or the dates of each run. The 91% is Anthropic’s own reported score, not an independent result. The slide’s curves are illustrative, as its footnote says, and we read it as a frame posted on X by Tanay Jaipuria. Anthropic and Carnegie Mellon were not contacted for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Anthropic’s S-1/A says Claude’s ExploitBench score rose from 40% with Opus 4.8 to 91% with Opus 5.5 in about four months, 51 points by our arithmetic. Its roadshow video places cyber second after coding, with its threshold crossed in April 2026.
A note on sourcing. Benchmark scores and filing quotations are from Anthropic’s S-1/A filed with the SEC on 10 November 2026, as reported by Anthropic. Video quotations are from Anthropic’s own captions file for its IPO roadshow video; slide text is read from a frame of the video posted on X by Tanay Jaipuria on 11 November 2026. Differences and ratios are our arithmetic. Anthropic was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
