{"schema_version":"onlylabs.public_signal.v1","title":"Anthropic Writing: Exploit Evals","description":"Anthropic writing signal with public source context, captured evidence pages, related signals, and data-business radar classification.","url":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48","json_url":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48/signal.json","generated_at":"2026-06-26T23:25:50.146Z","evidence_latest_fetched_at":"2026-06-17T20:04:15.799+00:00","signal_first_seen_at":"2026-06-17T20:00:31.646229+00:00","org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab","category_label":"Frontier lab","dossier_url":"https://onlylabs.fyi/labs/anthropic","dossier_json_url":"https://onlylabs.fyi/labs/anthropic/dossier.json"},"related_urls":{"signal":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48","signal_json":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48/signal.json","source":"https://www.anthropic.com/research/exploit-evals","lab_dossier":"https://onlylabs.fyi/labs/anthropic","lab_dossier_json":"https://onlylabs.fyi/labs/anthropic/dossier.json","analysis":"https://onlylabs.fyi/analysis/anthropic","analysis_json":"https://onlylabs.fyi/analysis/anthropic/analysis.json","analysis_evidence_json":"https://onlylabs.fyi/analysis/anthropic/evidence.json","category":"https://onlylabs.fyi/frontier","category_json":"https://onlylabs.fyi/frontier.json","category_feed":"https://onlylabs.fyi/frontier/feed.xml","category_signals_json":"https://onlylabs.fyi/signals.json","topic":"https://onlylabs.fyi/topics/talking","topic_signals_json":"https://onlylabs.fyi/topics/talking/signals.json","topic_feed":"https://onlylabs.fyi/topics/talking/feed.xml","data_business":{"radar":"https://onlylabs.fyi/data-radar","radar_json":"https://onlylabs.fyi/data-radar.json","opportunities":"https://onlylabs.fyi/opportunities","opportunities_json":"https://onlylabs.fyi/opportunities.json","lanes":[{"key":"evals","label":"Evals and quality","url":"https://onlylabs.fyi/data-radar/evals","json_url":"https://onlylabs.fyi/data-radar/evals/signals.json"}]}},"answer_pack":{"answer":"Anthropic published Exploit Evals. This talking signal gives public context for research themes, product direction, policy, or launch framing. High-signal details: Notable anthropic research post on evals. · EXPLOITBENCH: A CAPABILITY LADDER BENCHMARK FOR LLM CYBERSECURITY AGENTS A PREPRINT Seunghyun Lee Carnegie Mellon University Pittsburgh, PA 15213 David Brumley Carnegie.... onlylabs links this event to 3 captured evidence pages and 6 related writing signals. It also maps to Evals and quality in the data-business radar.","signal_desk":"talking","source_context":{"source_url":"https://www.anthropic.com/research/exploit-evals","source_host":"anthropic.com","occurred_at":"2026-05-22T00:00:00.000Z","first_seen_at":"2026-06-17T20:00:31.646229+00:00","date_source":"page.visible_date","context":null},"context_markers":[{"label":"Lab","value":"Anthropic","source":"signal"},{"label":"Signal desk","value":"talking","source":"signal"},{"label":"Source host","value":"anthropic.com","source":"source"},{"label":"PDF","value":"linked report","source":"source"},{"label":"Notability","value":"Notable anthropic research post on evals.","source":"signal"},{"label":"Radar lane","value":"Evals and quality","source":"radar"},{"label":"Matched term","value":"eval","source":"radar"},{"label":"Matched term","value":"evals","source":"radar"},{"label":"Watch term","value":"RL environments","source":"evidence"},{"label":"Watch term","value":"Eval methodology","source":"evidence"},{"label":"Watch term","value":"Infrastructure","source":"evidence"},{"label":"Watch term","value":"Safety and alignment","source":"evidence"},{"label":"Watch term","value":"Agents and tool use","source":"evidence"}],"evidence_coverage":{"target_pages":3,"captured_pages":3,"readable_pages":3,"capture_methods":["exa","plain"],"missing_page_urls":[],"failed_page_urls":[],"blocked_page_urls":[],"page_urls":["https://www.anthropic.com/research/exploit-evals","https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","https://exploitbench.ai/exploitbench.pdf"],"related_signals":6,"has_source_url":true,"latest_page_fetched_at":"2026-06-17T20:04:15.799+00:00"},"data_business":{"matches":true,"lanes":[{"key":"evals","label":"Evals and quality","url":"https://onlylabs.fyi/data-radar/evals","json_url":"https://onlylabs.fyi/data-radar/evals/signals.json"}],"matched_terms":["eval","evals"],"score":15,"reason":"Anthropic has a writing signal matching evals and quality."},"agent_handoff":{"signal_json":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48/signal.json","dossier_json":"https://onlylabs.fyi/labs/anthropic/dossier.json","analysis_json":"https://onlylabs.fyi/analysis/anthropic/analysis.json","analysis_evidence_json":"https://onlylabs.fyi/analysis/anthropic/evidence.json","topic_signals_json":"https://onlylabs.fyi/topics/talking/signals.json","topic_feed":"https://onlylabs.fyi/topics/talking/feed.xml","category_signals_json":"https://onlylabs.fyi/signals.json","data_radar_json":"https://onlylabs.fyi/data-radar.json","opportunities_json":"https://onlylabs.fyi/opportunities.json"},"analysis_playbook":{"objective":"Turn public writing and discussion into a readable map of research themes, product framing, policy posture, launch narratives, and market attention.","evidence_focus":["post title","source URL","captured page text","HN traction","linked model or paper references","publication date"],"extraction_questions":["Which themes are labs choosing to explain publicly?","Which posts are attracting outside discussion?","Which writing reframes a recent release, model, hiring wave, or policy stance?","Which posts mention data, evals, infrastructure, safety, or deployment workflows?"],"signal_questions":["What public theme, launch framing, or research direction does this writing signal expose?","Which themes are labs choosing to explain publicly?","Which posts are attracting outside discussion?","Which data-business lane explains this signal: Evals and quality?","Do the 6 related writing signals show a repeated pattern?"],"output_fields":["org","theme","public_framing","traction","data_business_lane","evidence_url"],"data_business_relevance":"Public writing supplies the narrative layer over raw signals and helps identify which frontier-lab priorities are becoming externally legible.","required_sources":[{"label":"signal_json","url":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48/signal.json","required":true},{"label":"source","url":"https://www.anthropic.com/research/exploit-evals","required":true},{"label":"dossier_json","url":"https://onlylabs.fyi/labs/anthropic/dossier.json","required":true},{"label":"analysis_evidence_json","url":"https://onlylabs.fyi/analysis/anthropic/evidence.json","required":true},{"label":"topic_signals_json","url":"https://onlylabs.fyi/topics/talking/signals.json","required":false},{"label":"data_radar_json","url":"https://onlylabs.fyi/data-radar.json","required":true}],"expected_output":["one-paragraph source-grounded interpretation","data-business implication","confidence and missing evidence","recommended next source to inspect"],"prompt_seed":"Using only the linked onlylabs JSON, captured source context, and cited evidence, analyze Anthropic's writing signal \"Exploit Evals\" for frontier lab strategy and data-business implications."},"semantic_triples":[{"subject":"Anthropic","predicate":"published","object":"Exploit Evals","text":"Anthropic published Exploit Evals."},{"subject":"Exploit Evals","predicate":"is classified as","object":"writing signal","text":"Exploit Evals is classified as writing signal."},{"subject":"Exploit Evals","predicate":"belongs to","object":"talking desk","text":"Exploit Evals belongs to talking desk."},{"subject":"Exploit Evals","predicate":"has evidence coverage","object":"3 captured evidence pages","text":"Exploit Evals has evidence coverage 3 captured evidence pages."},{"subject":"Exploit Evals","predicate":"matches data-business lanes","object":"Evals and quality","text":"Exploit Evals matches data-business lanes Evals and quality."},{"subject":"Exploit Evals","predicate":"has captured page count","object":"3","text":"Exploit Evals has captured page count 3."},{"subject":"Exploit Evals","predicate":"has readable page count","object":"3","text":"Exploit Evals has readable page count 3."},{"subject":"Exploit Evals","predicate":"has related signal count","object":"6","text":"Exploit Evals has related signal count 6."},{"subject":"Exploit Evals","predicate":"has analysis playbook objective","object":"Turn public writing and discussion into a readable map of research themes, product framing, policy posture, launch narratives, and market attention.","text":"Exploit Evals has analysis playbook objective Turn public writing and discussion into a readable map of research themes, product framing, policy posture, launch narratives, and market attention.."},{"subject":"Exploit Evals","predicate":"has source host","object":"anthropic.com","text":"Exploit Evals has source host anthropic.com."},{"subject":"Exploit Evals","predicate":"has lab","object":"Anthropic","text":"Exploit Evals has lab Anthropic."},{"subject":"Exploit Evals","predicate":"has signal desk","object":"talking","text":"Exploit Evals has signal desk talking."},{"subject":"Exploit Evals","predicate":"has source host","object":"anthropic.com","text":"Exploit Evals has source host anthropic.com."},{"subject":"Exploit Evals","predicate":"has pdf","object":"linked report","text":"Exploit Evals has pdf linked report."},{"subject":"Exploit Evals","predicate":"has notability","object":"Notable anthropic research post on evals.","text":"Exploit Evals has notability Notable anthropic research post on evals.."},{"subject":"Exploit Evals","predicate":"has radar lane","object":"Evals and quality","text":"Exploit Evals has radar lane Evals and quality."},{"subject":"Exploit Evals","predicate":"has matched term","object":"eval","text":"Exploit Evals has matched term eval."},{"subject":"Exploit Evals","predicate":"has matched term","object":"evals","text":"Exploit Evals has matched term evals."}]},"intelligence":{"signal_desk":"talking","answer":"Anthropic published Exploit Evals. This talking signal gives public context for research themes, product direction, policy, or launch framing. High-signal details: Notable anthropic research post on evals. · EXPLOITBENCH: A CAPABILITY LADDER BENCHMARK FOR LLM CYBERSECURITY AGENTS A PREPRINT Seunghyun Lee Carnegie Mellon University Pittsburgh, PA 15213 David Brumley Carnegie.... onlylabs links this event to 3 captured evidence pages and 6 related writing signals. It also maps to Evals and quality in the data-business radar.","semantic_triples":[{"subject":"Anthropic","predicate":"published","object":"Exploit Evals","text":"Anthropic published Exploit Evals."},{"subject":"Exploit Evals","predicate":"is classified as","object":"writing signal","text":"Exploit Evals is classified as writing signal."},{"subject":"Exploit Evals","predicate":"belongs to","object":"talking desk","text":"Exploit Evals belongs to talking desk."},{"subject":"Exploit Evals","predicate":"has evidence coverage","object":"3 captured evidence pages","text":"Exploit Evals has evidence coverage 3 captured evidence pages."},{"subject":"Exploit Evals","predicate":"matches data-business lanes","object":"Evals and quality","text":"Exploit Evals matches data-business lanes Evals and quality."}]},"signal":{"id":"b967bfd6-fc67-4b40-9a3e-7564ed60bf48","url":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48","json_url":"https://onlylabs.fyi/signals/b967bfd6-fc67-4b40-9a3e-7564ed60bf48/signal.json","source_url":"https://www.anthropic.com/research/exploit-evals","title":"Exploit Evals","summary":"Anthropic published a writing signal. onlylabs watches public writing for research themes, product direction, and model-launch context.","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2026-05-22T00:00:00.000Z","first_seen_at":"2026-06-17T20:00:31.646229+00:00","date_source":"page.visible_date","evidence_coverage":{"target_pages":3,"captured_pages":3,"readable_pages":3,"capture_methods":["exa","plain"],"missing_page_urls":[],"failed_page_urls":[],"blocked_page_urls":[],"page_urls":["https://www.anthropic.com/research/exploit-evals","https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","https://exploitbench.ai/exploitbench.pdf"]},"facets":{},"traction":{"github_stars":null,"hn_points":null,"hn_comments":null,"hn_story_id":null,"hf_downloads":null,"hf_likes":null},"data_radar":{"lanes":[{"key":"evals","label":"Evals and quality","url":"https://onlylabs.fyi/data-radar/evals"}],"score":15,"matched_terms":["eval","evals"],"reason":"Anthropic has a writing signal matching evals and quality."}},"primary_evidence_page":{"is_primary":true,"source_match":true,"url":"https://www.anthropic.com/research/exploit-evals","final_url":"https://www.anthropic.com/research/exploit-evals","title":"Exploit Evals","http_status":200,"content_type":"text/html; charset=utf-8","capture_method":"plain","fetched_at":"2026-06-17T20:04:03.492938+00:00","bytes":163062,"raw_path":"0fc33b995b409bd333129cf007f1979b04412dfda80055b25f6bf977baa0095c.html","content_hash":"4771dec5aef79c8a971240982375a9deca5cb24f942c0265bf53b559296b5d63","excerpt_chars":1200,"truncated":true,"excerpt":"Measuring LLMs’ ability to develop exploits \\ Anthropic Frontier Red Team Measuring LLMs’ ability to develop exploits May 22, 2026 Newton Cheng, Keane Lucas, Winnie Xiao, Nicholas Carlini, and Milad Nasr Introduction Claude Mythos Preview ’s ability to develop exploits is a step-change over previous frontier models. This was one of our primary motivations for rolling out the model carefully through Project Glasswing rather than through a general release. Mythos Preview is capable of finding complex vulnerabilities, but what concerned us most in our internal testing was that Mythos Preview could both turn vulnerabilities into exploit primitives, and combine those primitives together into complete end-to-end attack chains. When we published our Mythos Preview results , we measured its capabilities by having it search for novel zero-days and then build exploits for them. Qualitative evaluations like this are helpful for showcasing a model’s capabilities—but ideally, we would have high-quality quantitative benchmarks that let us measure them precisely. The problem we faced at the time we released Mythos Preview was that no existing public exploit benchmarks were difficult enough to..."},"evidence_pages":[{"is_primary":false,"source_match":false,"url":"https://exploitbench.ai/exploitbench.pdf","final_url":"https://exploitbench.ai/exploitbench.pdf","title":"Exploit Evals","http_status":200,"content_type":"application/pdf","capture_method":"exa","fetched_at":"2026-06-17T20:04:15.799+00:00","bytes":1461500,"raw_path":"0245e9f8dd929f641e912d224d9c23baee213943d9de7ac9f02f05e87a63c281.pdf","content_hash":"b78553e50bdb1f5c80f6bd4f6524657bf4cb37837d99759a1cb6447eef1bae4b","excerpt_chars":1200,"truncated":true,"excerpt":"EXPLOITBENCH: A CAPABILITY LADDER BENCHMARK FOR LLM CYBERSECURITY AGENTS A PREPRINT Seunghyun Lee Carnegie Mellon University Pittsburgh, PA 15213 David Brumley Carnegie Mellon University and Bugcrowd Pittsburgh, PA 15213 May 15, 2026 ABSTRACT Exploitation is not a binary event. It is a ladder of acquiring progressive capabilities, from executing a single buggy line of code to taking full control of the target. However, existing LLM security benchmarks treat a crash as exploitation success. That single binary outcome collapses the hard parts of exploitation: the transition from triggering a bug to constructing reusable primitives and control. As a result, we do not have an adequate understanding of where models stop when they do not succeed, or a way to measure their improvement trajectory. We present EXPLOITBENCH, a capability-graded benchmark that decomposes exploitation into 16 measurable flags, from coverage and crash through sandbox primitives, arbitrary read/write, control flow hijack, and arbitrary code execution. Each capability is verified by a deterministic oracle that uses a per-run randomized challenge-response for primitives, differential execution against ground truth..."},{"is_primary":false,"source_match":false,"url":"https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","final_url":"https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","title":"Anthropic Institute Agenda","http_status":200,"content_type":"application/pdf","capture_method":"plain","fetched_at":"2026-06-09T02:22:00.35164+00:00","bytes":23733776,"raw_path":"55b596ac8af2fe972d147b7a337236a0977b0661d555bffa93b6c190e54b7fc5.pdf","content_hash":"c33232e00848eb41f6f1faa3a72578ab887e74ce9e752aca954167e3645ad005","excerpt_chars":1200,"truncated":true,"excerpt":"System Card: Claude Mythos Preview April 7, 2026 anthropic.com Abstract This System Card describes Claude Mythos Preview, a large language model from Anthropic. Claude Mythos Preview is our most capable frontier model to date, and shows a striking leap in scores on many evaluation benchmarks compared to our previous frontier model, Claude Opus 4.6. This System Card assesses the model’s capabilities and reports many detailed safety evaluations. It covers tests relating to our Responsible Scaling Policy and our Frontier Compliance Framework, tests of cybersecurity skills, a wide-ranging alignment assessment, a model welfare assessment, and a new, largely qualitative section describing users’ experiences with the model. Claude Mythos Preview’s large increase in capabilities has led us to decide not to make it generally available. Instead, we are using it as part of a defensive cybersecurity program with a limited set of partners. The findings described in this System Card will be used to inform the release of future Claude models, as well as their associated safeguards. 2 Abstract 2 1 Introduction 9 1.1 Model training and characteristics 10 1.1.1 Training data and process 10 1.1.2..."}],"related_signals":[{"id":"9caf7d39-0d96-42d0-8b59-e72f7e502ec4","url":"https://onlylabs.fyi/signals/9caf7d39-0d96-42d0-8b59-e72f7e502ec4","source_url":"https://www.anthropic.com/research/economic-index-june-2026-report","title":"Economic Index June 2026 Report","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2026-06-26T15:40:57+00:00","first_seen_at":"2026-06-26T16:00:32.297045+00:00","date_source":"sitemap.lastmod"},{"id":"6c78c028-3ab4-4b33-86f7-d86c8ba9e3ba","url":"https://onlylabs.fyi/signals/6c78c028-3ab4-4b33-86f7-d86c8ba9e3ba","source_url":"https://www.anthropic.com/research/agents-in-biology","title":"Agents In Biology","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2026-06-08T00:00:00.000Z","first_seen_at":"2026-06-09T02:17:26.339488+00:00","date_source":"page.visible_date"},{"id":"e4fbfcdd-15b4-41b9-b011-fd83e7068ae9","url":"https://onlylabs.fyi/signals/e4fbfcdd-15b4-41b9-b011-fd83e7068ae9","source_url":"https://www.anthropic.com/research/making-claude-a-chemist","title":"Making Claude A Chemist","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2026-06-05T00:00:00.000Z","first_seen_at":"2026-06-09T02:17:26.339488+00:00","date_source":"page.visible_date"},{"id":"b8472bfd-c32b-4d72-ae98-d8a4f1a1278f","url":"https://onlylabs.fyi/signals/b8472bfd-c32b-4d72-ae98-d8a4f1a1278f","source_url":"https://www.anthropic.com/research/exploit","title":"Exploit","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2026-03-06T00:00:00.000Z","first_seen_at":"2026-06-17T20:00:31.646229+00:00","date_source":"page.visible_date"},{"id":"5f0f346c-0718-4d81-a174-d18cae51efa3","url":"https://onlylabs.fyi/signals/5f0f346c-0718-4d81-a174-d18cae51efa3","source_url":"https://www.anthropic.com/research/critical-infrastructure-defense","title":"Critical Infrastructure Defense","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2026-01-08T00:00:00.000Z","first_seen_at":"2026-06-17T20:00:31.646229+00:00","date_source":"page.visible_date"},{"id":"b8ad5d26-1905-4d23-a2e0-c2c477cbb462","url":"https://onlylabs.fyi/signals/b8ad5d26-1905-4d23-a2e0-c2c477cbb462","source_url":"https://www.anthropic.com/news/core-views-on-ai-safety","title":"Core Views On Ai Safety","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab"},"occurred_at":"2023-03-08T00:00:00.000Z","first_seen_at":"2026-06-09T02:17:26.339488+00:00","date_source":"page.visible_date"}]}