{
  "@context": "https://schema.org",
  "@id": "https://aggr.aymericbeaumet.com/items/simonwillison-net/2026-09-29-quoting-anthropic-frontier-red-team/#webpage",
  "@type": [
    "WebPage",
    "ArchiveComponent"
  ],
  "articleBody": "> We evaluate several models on 100 tasks from the \\[internal Binary Exploitation benchmark\\] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them.\n\n— [Anthropic Frontier Red Team](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities), GLM-5.3 and the spread of advanced cyber capabilities\n",
  "authors": [],
  "captured_at": "2026-09-29T23:14:03.129222144+00:00",
  "category": "programming",
  "content_html": "<blockquote>\n<p>We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them.</p>\n</blockquote>\n<p>— <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities\">Anthropic Frontier Red Team</a>, GLM-5.3 and the spread of advanced cyber capabilities</p>\n",
  "content_markdown": "> We evaluate several models on 100 tasks from the \\[internal Binary Exploitation benchmark\\] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them.\n\n— [Anthropic Frontier Red Team](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities), GLM-5.3 and the spread of advanced cyber capabilities\n",
  "dateCreated": "2026-09-29T23:14:03.129222144+00:00",
  "date_created": "2026-09-29T22:20:28+00:00",
  "date_modified": "2026-09-29T22:20:28+00:00",
  "date_published": "2026-09-29T22:20:28+00:00",
  "description": "We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although…",
  "external_url": "https://simonwillison.net/2026/Sep/29/anthropic-frontier-red-team/",
  "git": {
    "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e",
    "path": "items/simonwillison-net/2026/09/2026-09-29-quoting-anthropic-frontier-red-team.md",
    "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/simonwillison-net/2026/09/2026-09-29-quoting-anthropic-frontier-red-team.md"
  },
  "id": "urn:aggr:item:2f4493cca7ef5b603bfd4165a87cba98c0ad9d09",
  "isBasedOn": {
    "@id": "https://simonwillison.net/2026/Sep/29/anthropic-frontier-red-team/"
  },
  "mainEntity": {
    "@id": "https://simonwillison.net/2026/Sep/29/anthropic-frontier-red-team/",
    "@type": "CreativeWork",
    "archivedAt": {
      "@id": "https://aggr.aymericbeaumet.com/items/simonwillison-net/2026-09-29-quoting-anthropic-frontier-red-team/#webpage"
    },
    "dateModified": "2026-09-29T22:20:28+00:00",
    "datePublished": "2026-09-29T22:20:28+00:00",
    "headline": "Quoting Anthropic Frontier Red Team",
    "timeRequired": "PT1M",
    "url": "https://simonwillison.net/2026/Sep/29/anthropic-frontier-red-team/",
    "wordCount": 83
  },
  "name": "Archived snapshot: Quoting Anthropic Frontier Red Team",
  "preview": null,
  "reading_minutes": 1,
  "replicated_at": null,
  "source": "simonwillison-net",
  "source_name": "Simon Willison's Weblog",
  "summary": "We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them. — Anthropic Frontier Red Team, GLM-5.3 and the spread of advanced cyber capabilities Tags: anthropic, generative-ai, ai-security-research, glm, ai, ai-in-china, llms",
  "tags": [
    "ai",
    "ai-in-china",
    "ai-security-research",
    "anthropic",
    "generative-ai",
    "glm",
    "llms"
  ],
  "temporalCoverage": "2026-09-29T23:14:03.129222144+00:00",
  "title": "Quoting Anthropic Frontier Red Team",
  "url": "https://aggr.aymericbeaumet.com/items/simonwillison-net/2026-09-29-quoting-anthropic-frontier-red-team/",
  "word_count": 83
}