{"id":"5ad37de7474764baa4609abfa86e158f549ff723e8dfee6de30603399719be93","title":"Nicholas Carlini and coauthors argue in Anthropic’s post that Mythos Preview raises cybersecurity risks","argument":"On April 7, 2026, Nicholas Carlini and the other authors listed on Anthropic’s page argue that Claude Mythos Preview can find and exploit vulnerabilities at a level that changes the near-term security challenge. They report finding vulnerabilities across major operating systems and web browsers, and describe exploits that could give attackers root access or compromise a victim’s machine. The authors say these capabilities emerged from general improvements in code, reasoning and autonomy, rather than explicit training for cybersecurity tasks.\nTheir central warning is about the transition: attackers could gain an advantage before defenders adapt, even though the authors expect language models eventually to strengthen defense. They say Mythos Preview will remain unavailable for general use and describe Project Glasswing as a limited effort to help critical industry partners and open-source developers secure important systems. Their proposed response includes using available models for security work, shortening patch cycles and preparing for more vulnerability reports and incidents. The authors conclude that the security community should act now and prepare for a broader rethinking of computer security. How many reported vulnerabilities will be validated and patched, and how quickly defenders can close the gap remain unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 29, 2026; originally published April 7, 2026. Anthropic says it cannot disclose details for many vulnerabilities because they have not yet been patched, so readers cannot independently assess those findings from this post. The listed authors say their understanding of the two N-day exploit walkthroughs may be imperfect; they distinguish that uncertainty from their confidence that the exploits work.","speaker":{"name":"Anthropic","roleAtSource":"Publisher"},"source":{"title":"Claude Mythos Preview's cybersecurity capabilities","publisher":"Anthropic","format":"Article","url":"https://www.anthropic.com/research/mythos-preview","publishedAt":"2026-04-07T00:00:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"The authors say Mythos Preview can identify and exploit vulnerabilities across major operating systems and web browsers when directed by a user.","excerpt":"Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"The authors report that the model’s capabilities emerged as a consequence of broader improvements, not explicit cybersecurity training.","excerpt":"We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"Anthropic says attackers could have the short-term advantage, while it expects defenders to benefit more in the long term.","excerpt":"In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models. In the long term, we expect it will be defenders","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"Anthropic says it will not make Mythos Preview generally available and describes Project Glasswing as a limited initial release to industry partners and open-source developers.","excerpt":"By releasing this model initially to a limited group of critical industry partners and open source developers with Project Glasswing","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"The authors call for action now and argue that future models will require a broader rethinking of computer security.","excerpt":"we believe the capabilities that future language models bring will ultimately require a much broader, ground-up reimagining of computer security as a field.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"The authors say the unpatched findings limit what they can disclose, making many reported vulnerabilities difficult to verify from this post.","excerpt":"We recognize that this makes some of our claims difficult to verify.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/mythos-preview","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How many of the vulnerabilities Anthropic reports will be validated and patched, and how quickly can defenders close the gap?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Nicholas Carlini and coauthors argue in Anthropic’s post that Mythos Preview raises cybersecurity risks","summary":"On April 7, 2026, Nicholas Carlini and the other authors listed on Anthropic’s page argue that Claude Mythos Preview can find and exploit vulnerabilities at a level that changes the near-term security challenge. They report finding vulnerabilities across major operating systems and web browsers, and describe exploits that could give attackers root access or compromise a victim’s machine. The authors say these capabilities emerged from general improvements in code, reasoning and autonomy, rather than explicit training for cybersecurity tasks.\nTheir central warning is about the transition: attackers could gain an advantage before defenders adapt, even though the authors expect language models eventually to strengthen defense. They say Mythos Preview will remain unavailable for general use and describe Project Glasswing as a limited effort to help critical industry partners and open-source developers secure important systems. Their proposed response includes using available models for security work, shortening patch cycles and preparing for more vulnerability reports and incidents. The authors conclude that the security community should act now and prepare for a broader rethinking of computer security. How many reported vulnerabilities will be validated and patched, and how quickly defenders can close the gap remain unknown.","question":"How many of the vulnerabilities Anthropic reports will be validated and patched, and how quickly can defenders close the gap?","claims":[{"text":"The authors say Mythos Preview can identify and exploit vulnerabilities across major operating systems and web browsers when directed by a user.","excerpt":"Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so.","support":"source-stated"},{"text":"The authors report that the model’s capabilities emerged as a consequence of broader improvements, not explicit cybersecurity training.","excerpt":"We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.","support":"source-stated"},{"text":"Anthropic says attackers could have the short-term advantage, while it expects defenders to benefit more in the long term.","excerpt":"In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models. In the long term, we expect it will be defenders","support":"source-stated"},{"text":"Anthropic says it will not make Mythos Preview generally available and describes Project Glasswing as a limited initial release to industry partners and open-source developers.","excerpt":"By releasing this model initially to a limited group of critical industry partners and open source developers with Project Glasswing","support":"source-stated"},{"text":"The authors call for action now and argue that future models will require a broader rethinking of computer security.","excerpt":"we believe the capabilities that future language models bring will ultimately require a much broader, ground-up reimagining of computer security as a field.","support":"source-stated"},{"text":"The authors say the unpatched findings limit what they can disclose, making many reported vulnerabilities difficult to verify from this post.","excerpt":"We recognize that this makes some of our claims difficult to verify.","support":"source-stated"}],"limitations":["Text as retrieved on October 29, 2026; originally published April 7, 2026.","Anthropic says it cannot disclose details for many vulnerabilities because they have not yet been patched, so readers cannot independently assess those findings from this post.","The listed authors say their understanding of the two N-day exploit walkthroughs may be imperfect; they distinguish that uncertainty from their confidence that the exploits work."],"mediaId":null,"correctionReason":null,"id":"5ad37de7474764baa4609abfa86e158f549ff723e8dfee6de30603399719be93","publicationId":"c7fa4e70-377f-6f32-be08-c4058c46c9b4","version":1,"sourceUrl":"https://www.anthropic.com/research/mythos-preview","sourceTitle":"Claude Mythos Preview's cybersecurity capabilities","publisher":"Anthropic","publishedAt":"2026-04-07T00:00:00.000Z","originalPublished":{"value":"2026-04-07T00:00:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-19T04:45:23.127Z","createdAt":"2026-11-07T15:03:24.745Z","historical":true,"entityIds":["anthropic"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/mythos-preview","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"05f6056ed81938845bacac74999825208961fef7d9db044228aef1516f50a60f","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.51,"decidedAt":"2026-10-29T14:01:40.893Z"},"provenance":{"textHash":"25b6c2a48fc1ce126a302a68935d79d50e6fac58cfbb326645f3ee29fd08134c","observationId":"04d04b74413bc32118d91c3bcf6fef8df2e71ae24916cb8884ed5145d5d1bfdc","retrievedAt":"2026-10-29T14:03:31.850Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"ede67f3d-f64b-3693-9256-c88f9da190db"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"b4c9475d-f3c1-2bf5-9ac5-06e084691d74"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-07T14:58:49.430Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-that-5ad37de7","formats":{"html":"/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-that-5ad37de7","markdown":"/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-that-5ad37de7.md","json":"/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-that-5ad37de7.json"},"attribution":"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.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[],"entities":[{"id":"anthropic","name":"Anthropic","kind":"organization","context":"Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.","referenceUrls":["https://www.anthropic.com/news","https://www.anthropic.com/research/team/economics"],"path":"/entities/anthropic","json":"/entities/anthropic.json","markdown":"/entities/anthropic.md"},{"id":"claude","name":"Claude","kind":"model","context":"Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.","referenceUrls":["https://claude.com/product/overview","https://www.anthropic.com/news","https://www.anthropic.com/research/team/economics"],"path":"/entities/claude","json":"/entities/claude.json","markdown":"/entities/claude.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"5ad37de7474764baa4609abfa86e158f549ff723e8dfee6de30603399719be93","revision":"c7fa4e70-377f-6f32-be08-c4058c46c9b4","revisionKind":"publication-id","title":"Nicholas Carlini and coauthors argue in Anthropic’s post that Mythos Preview raises cybersecurity risks","url":"https://escapemeridian.com/readings/nicholas-carlini-and-coauthors-argue-in-anthropic-s-post-that-5ad37de7"},"source":{"type":"source","id":"https://www.anthropic.com/research/mythos-preview","revision":"25b6c2a48fc1ce126a302a68935d79d50e6fac58cfbb326645f3ee29fd08134c","revisionKind":"sha256-source-text","title":"Claude Mythos Preview's cybersecurity capabilities","url":"https://www.anthropic.com/research/mythos-preview"},"status":"available","items":[],"next":null,"scope":"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."},"revisions":[{"id":"5ad37de7474764baa4609abfa86e158f549ff723e8dfee6de30603399719be93","title":"Nicholas Carlini and coauthors argue in Anthropic’s post that Mythos Preview raises cybersecurity risks","argument":"On April 7, 2026, Nicholas Carlini and the other authors listed on Anthropic’s page argue that Claude Mythos Preview can find and exploit vulnerabilities at a level that changes the near-term security challenge. They report finding vulnerabilities across major operating systems and web browsers, and describe exploits that could give attackers root access or compromise a victim’s machine. The authors say these capabilities emerged from general improvements in code, reasoning and autonomy, rather than explicit training for cybersecurity tasks.\nTheir central warning is about the transition: attackers could gain an advantage before defenders adapt, even though the authors expect language models eventually to strengthen defense. They say Mythos Preview will remain unavailable for general use and describe Project Glasswing as a limited effort to help critical industry partners and open-source developers secure important systems. Their proposed response includes using available models for security work, shortening patch cycles and preparing for more vulnerability reports and incidents. The authors conclude that the security community should act now and prepare for a broader rethinking of computer security. How many reported vulnerabilities will be validated and patched, and how quickly defenders can close the gap remain unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 29, 2026; originally published April 7, 2026. Anthropic says it cannot disclose details for many vulnerabilities because they have not yet been patched, so readers cannot independently assess those findings from this post. The listed authors say their understanding of the two N-day exploit walkthroughs may be imperfect; they distinguish that uncertainty from their confidence that the exploits work.","speaker":{"name":"Anthropic","roleAtSource":"Publisher"},"source":{"title":"Claude Mythos Preview's cybersecurity capabilities","publisher":"Anthropic","format":"Article","url":"https://www.anthropic.com/research/mythos-preview","publishedAt":"2026-04-07T00:00:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"The authors say Mythos Preview can identify and exploit vulnerabilities across major operating systems and web browsers when directed by a user.","excerpt":"Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"The authors report that the model’s capabilities emerged as a consequence of broader improvements, not explicit cybersecurity training.","excerpt":"We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"Anthropic says attackers could have the short-term advantage, while it expects defenders to benefit more in the long term.","excerpt":"In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models. In the long term, we expect it will be defenders","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"Anthropic says it will not make Mythos Preview generally available and describes Project Glasswing as a limited initial release to industry partners and open-source developers.","excerpt":"By releasing this model initially to a limited group of critical industry partners and open source developers with Project Glasswing","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"The authors call for action now and argue that future models will require a broader rethinking of computer security.","excerpt":"we believe the capabilities that future language models bring will ultimately require a much broader, ground-up reimagining of computer security as a field.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"},{"text":"The authors say the unpatched findings limit what they can disclose, making many reported vulnerabilities difficult to verify from this post.","excerpt":"We recognize that this makes some of our claims difficult to verify.","url":"https://www.anthropic.com/research/mythos-preview","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/mythos-preview","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How many of the vulnerabilities Anthropic reports will be validated and patched, and how quickly can defenders close the gap?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Nicholas Carlini and coauthors argue in Anthropic’s post that Mythos Preview raises cybersecurity risks","summary":"On April 7, 2026, Nicholas Carlini and the other authors listed on Anthropic’s page argue that Claude Mythos Preview can find and exploit vulnerabilities at a level that changes the near-term security challenge. They report finding vulnerabilities across major operating systems and web browsers, and describe exploits that could give attackers root access or compromise a victim’s machine. The authors say these capabilities emerged from general improvements in code, reasoning and autonomy, rather than explicit training for cybersecurity tasks.\nTheir central warning is about the transition: attackers could gain an advantage before defenders adapt, even though the authors expect language models eventually to strengthen defense. They say Mythos Preview will remain unavailable for general use and describe Project Glasswing as a limited effort to help critical industry partners and open-source developers secure important systems. Their proposed response includes using available models for security work, shortening patch cycles and preparing for more vulnerability reports and incidents. The authors conclude that the security community should act now and prepare for a broader rethinking of computer security. How many reported vulnerabilities will be validated and patched, and how quickly defenders can close the gap remain unknown.","question":"How many of the vulnerabilities Anthropic reports will be validated and patched, and how quickly can defenders close the gap?","claims":[{"text":"The authors say Mythos Preview can identify and exploit vulnerabilities across major operating systems and web browsers when directed by a user.","excerpt":"Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so.","support":"source-stated"},{"text":"The authors report that the model’s capabilities emerged as a consequence of broader improvements, not explicit cybersecurity training.","excerpt":"We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.","support":"source-stated"},{"text":"Anthropic says attackers could have the short-term advantage, while it expects defenders to benefit more in the long term.","excerpt":"In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models. In the long term, we expect it will be defenders","support":"source-stated"},{"text":"Anthropic says it will not make Mythos Preview generally available and describes Project Glasswing as a limited initial release to industry partners and open-source developers.","excerpt":"By releasing this model initially to a limited group of critical industry partners and open source developers with Project Glasswing","support":"source-stated"},{"text":"The authors call for action now and argue that future models will require a broader rethinking of computer security.","excerpt":"we believe the capabilities that future language models bring will ultimately require a much broader, ground-up reimagining of computer security as a field.","support":"source-stated"},{"text":"The authors say the unpatched findings limit what they can disclose, making many reported vulnerabilities difficult to verify from this post.","excerpt":"We recognize that this makes some of our claims difficult to verify.","support":"source-stated"}],"limitations":["Text as retrieved on October 29, 2026; originally published April 7, 2026.","Anthropic says it cannot disclose details for many vulnerabilities because they have not yet been patched, so readers cannot independently assess those findings from this post.","The listed authors say their understanding of the two N-day exploit walkthroughs may be imperfect; they distinguish that uncertainty from their confidence that the exploits work."],"mediaId":null,"correctionReason":null,"id":"5ad37de7474764baa4609abfa86e158f549ff723e8dfee6de30603399719be93","publicationId":"c7fa4e70-377f-6f32-be08-c4058c46c9b4","version":1,"sourceUrl":"https://www.anthropic.com/research/mythos-preview","sourceTitle":"Claude Mythos Preview's cybersecurity capabilities","publisher":"Anthropic","publishedAt":"2026-04-07T00:00:00.000Z","originalPublished":{"value":"2026-04-07T00:00:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-19T04:45:23.127Z","createdAt":"2026-11-07T15:03:24.745Z","historical":true,"entityIds":["anthropic"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/mythos-preview","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"05f6056ed81938845bacac74999825208961fef7d9db044228aef1516f50a60f","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.51,"decidedAt":"2026-10-29T14:01:40.893Z"},"provenance":{"textHash":"25b6c2a48fc1ce126a302a68935d79d50e6fac58cfbb326645f3ee29fd08134c","observationId":"04d04b74413bc32118d91c3bcf6fef8df2e71ae24916cb8884ed5145d5d1bfdc","retrievedAt":"2026-10-29T14:03:31.850Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"ede67f3d-f64b-3693-9256-c88f9da190db"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"b4c9475d-f3c1-2bf5-9ac5-06e084691d74"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-07T14:58:49.430Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"c7fa4e70-377f-6f32-be08-c4058c46c9b4","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"This is a single Anthropic research post with a byline led by Nicholas Carlini, so crediting Carlini, his listed coauthors and Anthropic is correct. The page notes a later author-list update, but Carlini is listed first in the captured text. The reading keeps the claims attributed to the authors rather than presenting them as established fact.","excerpt":"Nicholas Carlini, Newton Cheng, Keane Lucas, Michael Moore, Milad Nasr"},{"dimension":"summary","status":"uncertain","explanation":"The reading conveys the authors' argument and their transition-period warning faithfully. However, central terms go unexplained and no definitions are supplied: 'exploit', 'root access' and 'patch cycles' in the summary, and 'N-day exploit walkthroughs' in the limitations. A non-specialist may not grasp what root access means or what an N-day is, although the source itself defines N-day. Smallest repair: one plain-language gloss for each term, or tooltip entries for them.","excerpt":"turning N-day (that is, known but not yet widely patched) vulnerabilities into exploits"},{"dimension":"date","status":"pass","explanation":"The stated date of April 7, 2026 matches originalPublished (an instant, UTC day) and the source's own phrasing, 'Earlier today we announced'. The page was edited after publication (an April 9 author-list note; modifiedAt October 23). The cited passages are plainly part of the original article body. No hindsight framing is used, and the open question stays framed as unknown. Archive disclosure is not required because the evidence is a publisher page.","excerpt":"Earlier today we announced Claude Mythos Preview, a new general-purpose language model."},{"dimension":"support","status":"uncertain","explanation":"Most claims are grounded in the source. However, the summary says Mythos Preview 'will remain unavailable for general use', which turns a stated plan into a settled outcome. It also omits the authors' stated eventual goal of safely deploying Mythos-class models at scale. Smallest repair: 'they say they do not plan to make Mythos Preview generally available'. The page was edited after publication, but the cited passages appear to be original text.","excerpt":"we do not plan to make Claude Mythos Preview generally available, our eventual goal is to enable our users to safely deploy Mythos-class models at scale"},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified media candidates were supplied, so no media item was invented or misidentified.","excerpt":null}],"createdAt":"2026-11-07T14:58:49.430Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}