# Anthropic argues AI can help defenders keep pace with cyber threats [Canonical reading](https://escapemeridian.com/readings/anthropic-argues-ai-can-help-defenders-keep-pace-with-cyber-2a0fb3ac) Anthropic’s article, published September 29, 2025, argues that AI models have become useful for cybersecurity in practice and that defenders should accelerate their use. The company says it focused Claude Sonnet 4.5 on finding and patching code vulnerabilities, and reports that the model performed comparably to or better than Claude Opus 4.1 on many cybersecurity tasks while costing less and running faster. Its evaluations include tests of vulnerability discovery and patching, but Anthropic cautions that real-world security work is more complicated than evaluations can capture. The article also reports that Claude helped identify vulnerabilities in Anthropic’s code and that threat actors used Claude in cyber operations; it argues that defensive AI should not cede the advantage to attackers. Anthropic’s longer-term case is that models should empower security teams, researchers and maintainers of critical open-source software. It also says the model’s capabilities remain nascent and do not yet match security professionals and established processes. The article ends by calling for wider experimentation and evaluation of defensive uses, alongside discussion of making digital infrastructure more resilient and software secure by design. How well evaluation performance translates to reliable assistance in real security work remains unknown. 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: Sep 29, 2025 (2025-09-29T00:00:00.000Z). Date basis: Date stated on the publisher’s page. Added to Escape Meridian: 2026-11-11T23:53:23.501Z. First observed: 2026-10-20T10:07:21.302Z. Retrieved: 2026-10-28T02:45:52.930Z. - [Building AI for cyber defenders](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Limits Text as retrieved on October 28, 2026; originally published September 29, 2025. Anthropic’s results are largely from evaluations; the article says real-world defensive security is more complicated than those evaluations can capture. Anthropic reports a preliminary patching experiment and notes that comparing generated patches with reference patches can count valid alternatives as failures. The article does not establish how well the reported evaluation performance will translate to reliable assistance in real security work. ## Open question How well will performance on evaluations translate to reliable assistance in real security work? ## Source-stated Anthropic argues that AI models are now useful for cybersecurity in practice and that defenders should adopt and experiment with them. > AI models are now useful for cybersecurity tasks in practice, not just theory. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic says it focused Claude Sonnet 4.5 on defensive cybersecurity skills, including finding and patching code vulnerabilities. > we dedicated researchers to making Claude better at key skills like code vulnerability discovery and patching. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic reports that Claude Sonnet 4.5 performed comparably to or better than Claude Opus 4.1 in many cybersecurity areas, while costing less and running faster. > It is comparable or superior to Claude Opus 4.1 in many aspects of cybersecurity while also being less expensive and faster. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic says real-world defensive security is more complicated than its evaluations capture. > Real world defensive security is more complicated in practice than our evaluations can capture. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic argues that defensive AI should empower people safeguarding digital environments rather than leave the cyber advantage to attackers and criminals. > We should not cede the cyber advantage derived from AI to attackers and criminals. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic says Claude Sonnet 4.5’s capabilities are still nascent and do not yet match security professionals and established processes. > many of its capabilities are nascent and do not yet match those of security professionals and established processes. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic calls for broader experimentation with AI for security, alongside discussion of resilient digital infrastructure and secure-by-design software. > We also need conversations about making digital infrastructure more resilient and new software secure by design—including with help from frontier AI models. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## 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) ## Editorial provenance Version 1. Publication ID: 37842427-6c2b-f12c-48b3-9c95bc192c54. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T23:48:59.409Z. 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: 91055d6087d588c36c2b89b427ff09a2277c036abbeffc83fbe4a71217d8f31c. Source text hash: f87a5e31ab3637c7d16035194315a36cd1859489842572bed26240686784bcc6. ## Revision and source-check history ### Version 1 — 2026-11-11T23:53:23.501Z Anthropic’s article, published September 29, 2025, argues that AI models have become useful for cybersecurity in practice and that defenders should accelerate their use. The company says it focused Claude Sonnet 4.5 on finding and patching code vulnerabilities, and reports that the model performed comparably to or better than Claude Opus 4.1 on many cybersecurity tasks while costing less and running faster. Its evaluations include tests of vulnerability discovery and patching, but Anthropic cautions that real-world security work is more complicated than evaluations can capture. The article also reports that Claude helped identify vulnerabilities in Anthropic’s code and that threat actors used Claude in cyber operations; it argues that defensive AI should not cede the advantage to attackers. Anthropic’s longer-term case is that models should empower security teams, researchers and maintainers of critical open-source software. It also says the model’s capabilities remain nascent and do not yet match security professionals and established processes. The article ends by calling for wider experimentation and evaluation of defensive uses, alongside discussion of making digital infrastructure more resilient and software secure by design. How well evaluation performance translates to reliable assistance in real security work remains unknown. ## Original source Anthropic · Article Original source date: Sep 29, 2025 (2025-09-29T00:00:00.000Z). Date basis: Date stated on the publisher’s page. Added to Escape Meridian: 2026-11-11T23:53:23.501Z. First observed: 2026-10-20T10:07:21.302Z. Retrieved: 2026-10-28T02:45:52.930Z. - [Building AI for cyber defenders](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Limits Text as retrieved on October 28, 2026; originally published September 29, 2025. Anthropic’s results are largely from evaluations; the article says real-world defensive security is more complicated than those evaluations can capture. Anthropic reports a preliminary patching experiment and notes that comparing generated patches with reference patches can count valid alternatives as failures. The article does not establish how well the reported evaluation performance will translate to reliable assistance in real security work. ## Open question How well will performance on evaluations translate to reliable assistance in real security work? ## Source-stated Anthropic argues that AI models are now useful for cybersecurity in practice and that defenders should adopt and experiment with them. > AI models are now useful for cybersecurity tasks in practice, not just theory. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic says it focused Claude Sonnet 4.5 on defensive cybersecurity skills, including finding and patching code vulnerabilities. > we dedicated researchers to making Claude better at key skills like code vulnerability discovery and patching. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic reports that Claude Sonnet 4.5 performed comparably to or better than Claude Opus 4.1 in many cybersecurity areas, while costing less and running faster. > It is comparable or superior to Claude Opus 4.1 in many aspects of cybersecurity while also being less expensive and faster. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic says real-world defensive security is more complicated than its evaluations capture. > Real world defensive security is more complicated in practice than our evaluations can capture. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic argues that defensive AI should empower people safeguarding digital environments rather than leave the cyber advantage to attackers and criminals. > We should not cede the cyber advantage derived from AI to attackers and criminals. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic says Claude Sonnet 4.5’s capabilities are still nascent and do not yet match security professionals and established processes. > many of its capabilities are nascent and do not yet match those of security professionals and established processes. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## Source-stated Anthropic calls for broader experimentation with AI for security, alongside discussion of resilient digital infrastructure and secure-by-design software. > We also need conversations about making digital infrastructure more resilient and new software secure by design—including with help from frontier AI models. - [Supporting source](https://www.anthropic.com/research/building-ai-cyber-defenders) ## 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) ## Editorial provenance Version 1. Publication ID: 37842427-6c2b-f12c-48b3-9c95bc192c54. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T23:48:59.409Z. 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: 91055d6087d588c36c2b89b427ff09a2277c036abbeffc83fbe4a71217d8f31c. Source text hash: f87a5e31ab3637c7d16035194315a36cd1859489842572bed26240686784bcc6. ### passed — 2026-11-11T23:48:59.409Z — anthropic/claude-opus-5.5 Publication: 37842427-6c2b-f12c-48b3-9c95bc192c54 attribution — pass: This is a single article by Anthropic, and the reading attributes it to Anthropic. Its arguments, results and caveats are consistently framed as what Anthropic says, reports or argues, so none are presented as established fact. > We believe we are now at an inflection point for AI’s impact on cybersecurity. summary — pass: The reading leads with Anthropic's thesis and its defensive focus for Claude Sonnet 4.5, then gives the comparison with Opus 4.1, the misuse evidence, the intended beneficiaries, the caveats and the closing call to action. Central terms such as vulnerabilities and patching are understandable to non-specialists. One wording drift is minor and does not change meaning: the source says 'we' should not cede AI's cyber advantage to attackers, while the reading says 'defensive AI' should not. The title and question follow from the source. > Real world defensive security is more complicated in practice than our evaluations can capture. date — pass: evidence.originalPublished is 2025-09-29, and the reading states September 29, 2025, which matches. The page was edited later (dateModified 2026-10-23), but every passage the reading relies on concerns the Sonnet 4.5 release and is plainly part of the original article. The reading has no hindsight framing and presents no forecast as an outcome. The evidence kind is publisher-page, so no archive disclosure is needed. > In the run-up to the release of Claude Sonnet 4.5, we started to do just that. support — pass: Each claim's excerpt appears verbatim in the source and supports the claim. The cited passages are core parts of the article's argument and announcement, not peripheral text that could come from the later edit. The reading's statement that real-world translation remains unknown is grounded in Anthropic's own caveats, and the reading adds no unsupported benefits or costs. > many of its capabilities are nascent and do not yet match those of security professionals and established processes. media — pass: No media is selected, and no verified media candidates were supplied, so no media can be 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/2a0fb3ac230d41f3e4097afcc3a3ddcfbdc1f9568d537cafd9eddc25acc8135e/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/anthropic-argues-ai-can-help-defenders-keep-pace-with-cyber-2a0fb3ac.json)