# Nicholas Carlini and coauthors argue in Anthropic’s post for faster vulnerability defence [Canonical reading](https://escapemeridian.com/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-for-238159b3) Nicholas Carlini and coauthors published this post on February fifth, twenty twenty-six, arguing that Claude Opus 4.6 can find high-severity vulnerabilities in well-tested open-source code without specialised setup. Unlike fuzzers, which test code with many inputs, Claude can inspect code and past fixes to look for similar bugs that remain. The authors report that it found vulnerabilities in projects tested for years, and say they validated findings before reporting them. Human researchers validated the vulnerabilities and wrote initial patches; external researchers helped as the work grew. The authors conclude that language models can add value alongside existing discovery tools, while creating a dual-use risk: they can aid defence or misuse. Anthropic describes new detection and enforcement measures, including possible real-time blocking of traffic it detects as malicious, while acknowledging this could hinder legitimate research and defensive work. The authors warn that industry-standard 90-day disclosure windows may not keep pace with the speed and volume of findings. They also forecast that models may soon exceed expert human researchers in speed and scale. How disclosure practices and misuse controls can keep pace without obstructing legitimate security work remains unresolved. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source Anthropic · Article Original source date: Feb 5, 2026 (2026-02-05T00:00:00.000Z). Date basis: Date stated on the publisher’s page. Added to Escape Meridian: 2026-11-08T09:23:24.073Z. First observed: 2026-10-18T23:37:20.787Z. Retrieved: 2026-10-30T03:11:48.304Z. - [LLM-discovered 0 days](https://www.anthropic.com/research/zero-days) ## Limits Text as retrieved on October 30, 2026; originally published February 5, 2026. The authors focused on memory corruption vulnerabilities, which they say are easier to validate than logic errors; the post does not establish how well the approach finds other vulnerability types. ## Open question How can vulnerability disclosure and misuse controls keep pace with faster discovery without obstructing legitimate security research and defensive work? ## Source-stated The authors argue that Claude Opus 4.6 can find high-severity vulnerabilities in well-tested code without specialised scaffolding. > Opus 4.6 found high-severity vulnerabilities, some that had gone undetected for decades. - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source-stated Anthropic says its team validated reported bugs and used human security researchers to validate vulnerabilities and write initial patches. > our own security researchers validated each vulnerability and wrote patches by hand. - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source-stated The authors say proposed real-time intervention could create friction for legitimate research and defensive work. > This will create friction for legitimate research and some defensive work - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source-stated The authors warn that industry-standard disclosure windows may not keep pace with the speed and volume of language-model-discovered bugs. > Industry-standard 90-day windows may not hold up against the speed and volume of LLM-discovered bugs - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source and contributor context Anthropic · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: c7a76c40-8fae-feb6-95c2-59fcb9be0248. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T09:18:41.625Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: cce2e55c38690be41d4c1fd37299735755fb64a78965ac1bdb01474de6fe5e43. Source text hash: 229eb9e1bcbe494a73880c0817b2ad73624146582fbf92733c77223515f74e02. ## Revision and source-check history ### Version 1 — 2026-11-08T09:23:24.073Z Nicholas Carlini and coauthors published this post on February fifth, twenty twenty-six, arguing that Claude Opus 4.6 can find high-severity vulnerabilities in well-tested open-source code without specialised setup. Unlike fuzzers, which test code with many inputs, Claude can inspect code and past fixes to look for similar bugs that remain. The authors report that it found vulnerabilities in projects tested for years, and say they validated findings before reporting them. Human researchers validated the vulnerabilities and wrote initial patches; external researchers helped as the work grew. The authors conclude that language models can add value alongside existing discovery tools, while creating a dual-use risk: they can aid defence or misuse. Anthropic describes new detection and enforcement measures, including possible real-time blocking of traffic it detects as malicious, while acknowledging this could hinder legitimate research and defensive work. The authors warn that industry-standard 90-day disclosure windows may not keep pace with the speed and volume of findings. They also forecast that models may soon exceed expert human researchers in speed and scale. How disclosure practices and misuse controls can keep pace without obstructing legitimate security work remains unresolved. ## Original source Anthropic · Article Original source date: Feb 5, 2026 (2026-02-05T00:00:00.000Z). Date basis: Date stated on the publisher’s page. Added to Escape Meridian: 2026-11-08T09:23:24.073Z. First observed: 2026-10-18T23:37:20.787Z. Retrieved: 2026-10-30T03:11:48.304Z. - [LLM-discovered 0 days](https://www.anthropic.com/research/zero-days) ## Limits Text as retrieved on October 30, 2026; originally published February 5, 2026. The authors focused on memory corruption vulnerabilities, which they say are easier to validate than logic errors; the post does not establish how well the approach finds other vulnerability types. ## Open question How can vulnerability disclosure and misuse controls keep pace with faster discovery without obstructing legitimate security research and defensive work? ## Source-stated The authors argue that Claude Opus 4.6 can find high-severity vulnerabilities in well-tested code without specialised scaffolding. > Opus 4.6 found high-severity vulnerabilities, some that had gone undetected for decades. - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source-stated Anthropic says its team validated reported bugs and used human security researchers to validate vulnerabilities and write initial patches. > our own security researchers validated each vulnerability and wrote patches by hand. - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source-stated The authors say proposed real-time intervention could create friction for legitimate research and defensive work. > This will create friction for legitimate research and some defensive work - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source-stated The authors warn that industry-standard disclosure windows may not keep pace with the speed and volume of language-model-discovered bugs. > Industry-standard 90-day windows may not hold up against the speed and volume of LLM-discovered bugs - [Supporting source](https://www.anthropic.com/research/zero-days) ## Source and contributor context Anthropic · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: c7a76c40-8fae-feb6-95c2-59fcb9be0248. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T09:18:41.625Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: cce2e55c38690be41d4c1fd37299735755fb64a78965ac1bdb01474de6fe5e43. Source text hash: 229eb9e1bcbe494a73880c0817b2ad73624146582fbf92733c77223515f74e02. ### passed — 2026-11-08T09:18:41.625Z — anthropic/claude-opus-5.5 Publication: c7a76c40-8fae-feb6-95c2-59fcb9be0248 attribution — pass: The reading attributes the post to Nicholas Carlini and coauthors at Anthropic, matching the byline, and frames claims as the authors' or Anthropic's statements. The page notes a February 6 edit that updated the author list. Naming the first listed author with coauthors stays consistent with that list. > Nicholas Carlini\*, Keane Lucas\*, Evyatar Ben Asher\*, Newton Cheng, Hasnain Lakhani, David Forsythe, and Kyla Guru summary — pass: The summary leads with the authors' central argument and explains fuzzers and dual-use risk in plain terms. It covers validation, safeguards and their admitted friction, the disclosure-window warning and the attributed forecast. The question follows from the source's own concerns. The 500-vulnerability figure is omitted, but its absence does not distort the reading's meaning. > Our view is this is a moment to move quickly—to empower defenders and secure as much code as possible while the window exists. date — pass: February 5, 2026 matches originalPublished at day precision. The kind is publisher-page, so no archive disclosure is needed. The forecast about exceeding human researchers is reported as the authors' forecast. The reading does not rely on any passage from the later page modification (dateModified October 2026). The only in-text edit note concerns the author list. > Edited February 6, 2026: Updated the author list support — pass: Each claim is supported by passages from the main article body. These cover out-of-the-box discovery, human validation and patching, admitted friction from real-time intervention, and the 90-day disclosure warning. The memory-corruption limitation is also source-supported. The cited body text appears to be original article content rather than a later edit. Source-grounding here does not independently verify the real-world claims. > Claude Opus 4.6 can find meaningful 0-day vulnerabilities in well-tested codebases, even without specialized scaffolding. media — pass: No media is selected, and no verified media candidates were supplied, so nothing is misidentified or irrelevant. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/238159b32b50f0f3b2a6201333ea42f59c0900c40b9806eec462c470881359c3/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-for-238159b3.json)