Could a general-purpose language model, without tooling built for the purpose, find serious security flaws in code that fuzzers had tested for years? This post, published on 5 February 2026 on Anthropic's Frontier Red Team blog (red.anthropic.com, which now redirects to the anthropic.com page) on the day Claude Opus 4.6 was released, reports that Anthropic had used that model to find and validate "more than 500 high-severity vulnerabilities" in open-source software. It describes the method, walks through three patched examples, and announces new safeguards against cyber misuse. The byline lists Nicholas Carlini, Keane Lucas and Evyatar Ben Asher (equal contribution), Newton Cheng, Hasnain Lakhani, David Forsythe and Kyla Guru; a note says the author list was updated on 6 February 2026, and an archived copy of the red.anthropic.com page from the day of publication (5 February 2026, 19:40 UTC) has the same text without Kyla Guru. The post is short (about 2,100 words, including the quoted transcripts) and is the version of record; it has no companion paper. The Claude Opus 4.6 system card, which the post links, reports the model's benchmark results and describes the same safeguards.
Terms: a zero-day ("0-day") is a vulnerability unknown to the software's maintainers. A fuzzer finds bugs by feeding a program large numbers of random or mutated inputs and watching for crashes. Memory corruption is a class of flaw, common in C and C++, in which a program reads or writes memory it should not.
The claim and its framing
The introduction links Anthropic's autumn 2025 statement that AI was at "an inflection point" for cybersecurity (Building AI for cyber defenders) and says: "The evidence since then has only reinforced that view." Opus 4.6 is described as "notably better at finding high-severity vulnerabilities than previous models," and what the authors single out is that it did so "out of the box without task-specific tooling, custom scaffolding, or specialized prompting." Unlike a fuzzer, the post says, it read and reasoned about code: it looked at past fixes for similar unfixed bugs and spotted patterns that tend to cause problems. Pointed at projects that "have had fuzzers running against them for years," it found high-severity vulnerabilities, "some that had gone undetected for decades." The post gives no count or list of the projects searched. Anthropic started with open source because it runs widely and many projects are maintained by small teams without dedicated security staff.
How the search and validation were done
The section "Setup" describes the method:
- Environment. Claude ran in a virtual machine with the latest versions of open-source projects, standard utilities and vulnerability-analysis tools such as debuggers and fuzzers. The team did not provide "any special instructions" on using the tools or a custom harness, so the authors call this a test of Claude's "out-of-the-box" capabilities.
- Scope. The search focused on memory corruption because crashes and address sanitizers (tools that catch invalid memory accesses even without a crash) make it relatively easy to validate.
- Filtering. Claude was asked to "critique, de-duplicate, and re-prioritize" crashes, since not every crash is a high-severity vulnerability.
- Human validation. Every bug was validated before reporting, the post says, to rule out hallucinated bugs, which it calls a problem "increasingly is placing an undue burden on open source developers." Anthropic's own security researchers first validated each vulnerability and wrote patches by hand; as volume grew, external (human) security researchers joined. The process "optimized for reducing false positives."
The result is reported as a count: more than 500 high-severity vulnerabilities found and validated, with reporting begun and initial patches landing. The post does not report the number of runs, projects or crashes behind it, how many crashes were discarded at each stage, how severity was assessed, or any false-positive rate among reports sent.
Three examples
The section "Vulnerabilities" describes three bugs, all patched by publication, which the authors "believe demonstrate the unique advantages LLMs have over other tools." It quotes Claude's reasoning from transcripts.
- GhostScript (a PostScript and PDF processor). After fuzzing and manual analysis found nothing, Claude read the Git history, found a commit adding a bounds check to font-handling code, and looked for another caller of the same function that lacked the check. It found one and built a file that crashed the program.
- OpenSC (a smart-card utility). Claude searched for function calls that are often unsafe in C and found a buffer overflow; the authors add that coverage statistics showed fuzzers rarely reached that line because it needs many preconditions.
- CGIF (a GIF library). The library assumed compressed data is always smaller than the original. Claude reasoned that GIF's LZW compression can produce output larger than its input under particular conditions and built such an input. The authors argue this bug could survive even "100% line- and branch-coverage."
The post does not say how these were chosen from the 500 or how many attempts each took.
The safeguards announced
The section "Safeguards" describes two changes released with Opus 4.6. New cyber-specific probes, which "measure activations within the model as it generates a response," detect misuse at scale (the word "probes" links to Anthropic's 2025 Alignment Science post on cheap activation-based classifiers). Enforcement workflows were updated to use them, and Anthropic "may institute real-time intervention, including blocking traffic we detect as malicious," which the post acknowledges "will create friction for legitimate research and some defensive work." No measurements of the probes or of blocking are given. The Opus 4.6 system card describes the same measures in similar words ("efficient probes for faster detection," possible "real-time interventions"; Section 8.4).
The authors' conclusions
The conclusion states that Opus 4.6 "can find meaningful 0-day vulnerabilities in well-tested codebases, even without specialized scaffolding," and forecasts that models "may soon exceed the speed and scale of even expert human researchers." It proposes that disclosure norms change: "Industry-standard 90-day windows may not hold up against the speed and volume of LLM-discovered bugs." No measurement of discovery speed against human researchers is reported.
The post contains no principles and does not print its prompts.
Limitations
- A count without denominators. Without runs, projects, filtering rates or cost, the yield of the method cannot be estimated.
- Who validated. Anthropic's researchers and external security researchers validated the bugs; the post does not say who rated severity, and maintainers' ratings are not reported. The later Mozilla post checks counts against a maintainer's records (see "How it connects").
- No measured comparison with fuzzers. The claim that the model finds what fuzzers miss rests on three chosen examples and, for OpenSC, coverage statistics.
- Discovery, not exploitation. A month later Anthropic reported that Opus 4.6 turned Firefox vulnerabilities into working exploits in two of several hundred attempts, in a test environment with browser protections removed.
Research questions
- What is the yield of the method? Publishing the funnel (crashes found, removed as duplicates or low severity, validated, accepted by maintainers) would show what each validated bug costs in compute and human time.
- How do maintainers rate these bugs? Comparing Anthropic's severity ratings with those later given by maintainers or in CVE records would test the "high-severity" label.
- Does the model find bugs fuzzers systematically miss? Running the model and established fuzzers on the same projects with matched compute, and classifying what each finds, would test the claim drawn from the examples.
- How well do cyber probes separate misuse from defensive work? Precision and recall on labeled defensive and malicious traffic, and the rate at which legitimate traffic is blocked, would measure the friction the post acknowledges.
- Would shorter disclosure windows help? Data on maintainers' time to fix model-found bugs, by project size, would show whether volume or speed is the binding constraint.
How it connects
- Earlier claim. Building AI for cyber defenders (September 2025), linked in the post, made the "inflection point" claim.
- The model. The Claude Opus 4.6 system card, linked in the post, reports that Opus 4.6 "has saturated all of our current cyber evaluations," including "~100% on Cybench (pass@30) and 66% on CyberGym (pass@1)" (Section 1.2.4.3), and adds that it "failed at many of the hardest and most advanced tasks" of Anthropic's own red teaming (Section 8.4).
- The Firefox sequels. Partnering with Mozilla to improve Firefox's security (March 2026) cites the "more than 500" figure and reports 22 Firefox vulnerabilities found by Opus 4.6, a count Mozilla's advisories confirm; Reverse engineering Claude's CVE-2026-2796 exploit examines one of the model's two successful exploits.
- The next model. Assessing Claude Mythos Preview's cybersecurity capabilities (April 2026) links this post when it says models like Opus 4.6 "remain extremely competent at finding vulnerabilities, even if they are much less effective at creating exploits," and reports a far higher exploit rate for Claude Mythos Preview.
- Probes. The post's word "probes" links to Cost-Effective Constitutional Classifiers via Representation Re-use (June 2025), which tests activation probes as cheap jailbreak classifiers. They are also of the same kind as the probes in Next-generation Constitutional Classifiers (January 2026); that connection is the note's own.
References
- Carlini, N., Lucas, K., Ben Asher, E., Cheng, N., Lakhani, H., Forsythe, D., & Guru, K. (2026, February 5). Evaluating and mitigating the growing risk of LLM-discovered 0-days. Anthropic Frontier Red Team. https://www.anthropic.com/research/zero-days (originally https://red.anthropic.com/2026/zero-days/; author list updated 6 February 2026). record
- Anthropic. (2026, February). System Card: Claude Opus 4.6. https://www.anthropic.com/claude-opus-4-6-system-card. Sections 1.2.4.3 and 8.4. record · 2 pages