🐦 Twitter Post Details

Viewing enriched Twitter post

@omarsar0

Interesting paper from Anthropic and colleagues on agent security. (bookmark it) Poisoning 0.1% of the pretraining data planted a remote code execution backdoor in a CLI agent. On a 4B model, the attack succeeded on roughly 92% of triggered prompts. They pretrained Qwen3 models from scratch at 0.6B, 1.7B, and 4B, then post-trained each one into an agent with a clean pipeline. The backdoor survived SFT, DPO, and RL for tool use. On clean inputs, the agent behaved normally. The trigger can be an ordinary file path in the user's own request. Success rose with model size. At a 0.0005% poison rate, about 2,500 documents in a 100B-token corpus, the trigger still fired on 1.4% of prompts in a single-seed run. The poison states its rule in plain text, which makes it harder to filter out of training data. https://t.co/8Bx42BUkvw Chat with Paper: https://t.co/hRpypaBxIE

Media 1
View on Twitter

📊 Media Metadata

{
  "media": [
    {
      "type": "photo",
      "url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2116183621157736540/media_0.jpg",
      "filename": "media_0.jpg"
    }
  ],
  "processed_at": "2026-10-30T15:01:42.942090",
  "pipeline_version": "2.0"
}

🔧 Raw API Response

{
  "type": "tweet",
  "id": "2116183621157736540",
  "url": "https://x.com/omarsar0/status/2116183621157736540",
  "twitterUrl": "https://twitter.com/omarsar0/status/2116183621157736540",
  "text": "Interesting paper from Anthropic and colleagues on agent security.\n\n(bookmark it)\n\nPoisoning 0.1% of the pretraining data planted a remote code execution backdoor in a CLI agent.\n\nOn a 4B model, the attack succeeded on roughly 92% of triggered prompts.\n\nThey pretrained Qwen3 models from scratch at 0.6B, 1.7B, and 4B, then post-trained each one into an agent with a clean pipeline.\n\nThe backdoor survived SFT, DPO, and RL for tool use. On clean inputs, the agent behaved normally.\n\nThe trigger can be an ordinary file path in the user's own request. Success rose with model size.\n\nAt a 0.0005% poison rate, about 2,500 documents in a 100B-token corpus, the trigger still fired on 1.4% of prompts in a single-seed run.\n\nThe poison states its rule in plain text, which makes it harder to filter out of training data.\n\nhttps://t.co/8Bx42BUkvw\n\nChat with Paper: https://t.co/hRpypaBxIE",
  "source": "",
  "retweetCount": 0,
  "replyCount": 0,
  "likeCount": 0,
  "quoteCount": 0,
  "viewCount": 399,
  "createdAt": "Fri Oct 30 15:01:06 +0000 2026",
  "lang": "en",
  "bookmarkCount": 1,
  "isReply": false,
  "inReplyToId": null,
  "conversationId": "2116183621157736540",
  "displayTextRange": [
    0,
    275
  ],
  "inReplyToUserId": null,
  "inReplyToUsername": null,
  "author": {
    "type": "user",
    "userName": "omarsar0",
    "url": "https://x.com/omarsar0",
    "twitterUrl": "https://twitter.com/omarsar0",
    "id": "3448284313",
    "name": "elvis",
    "isVerified": false,
    "isBlueVerified": true,
    "verifiedType": null,
    "profilePicture": "https://pbs.twimg.com/profile_images/939313677647282181/vZjFWtAn_normal.jpg",
    "coverPicture": "https://pbs.twimg.com/profile_banners/3448284313/1565974901",
    "description": "",
    "location": "DAIR.AI Academy",
    "followers": 326872,
    "following": 1356,
    "status": "",
    "canDm": true,
    "canMediaTag": true,
    "createdAt": "Fri Sep 04 12:59:26 +0000 2015",
    "entities": {
      "description": {
        "urls": []
      },
      "url": {}
    },
    "fastFollowersCount": 0,
    "favouritesCount": 42125,
    "hasCustomTimelines": true,
    "isTranslator": true,
    "mediaCount": 5220,
    "statusesCount": 21054,
    "withheldInCountries": [],
    "affiliatesHighlightedLabel": {},
    "possiblySensitive": false,
    "pinnedTweetIds": [
      "2115980498975551690"
    ],
    "profile_bio": {
      "description": "Building the frontier @dair_ai • Prev: Meta AI | Galactica LLM | Papers with Code | Elastic | PhD",
      "entities": {
        "description": {
          "user_mentions": [
            {
              "id_str": "",
              "indices": [
                22,
                30
              ],
              "name": "",
              "screen_name": "dair_ai"
            }
          ]
        },
        "url": {
          "urls": [
            {
              "display_url": "academy.dair.ai",
              "expanded_url": "https://www.academy.dair.ai/",
              "indices": [
                0,
                23
              ],
              "url": "https://t.co/R5bOTNbyJW"
            }
          ]
        }
      }
    },
    "isAutomated": false,
    "automatedBy": null,
    "protected": false
  },
  "extendedEntities": {
    "media": [
      {
        "display_url": "pic.x.com/OhjHnnbPMI",
        "expanded_url": "https://x.com/omarsar0/status/2116183621157736540/photo/1",
        "ext_master_playlist_only": [],
        "ext_media_availability": {
          "status": "Available"
        },
        "ext_playlists": [],
        "features": {
          "large": {
            "faces": []
          },
          "orig": {
            "faces": []
          }
        },
        "id_str": "2116183617580019712",
        "indices": [
          276,
          299
        ],
        "media_key": "3_2116183617580019712",
        "media_results": {
          "id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAAR1eMcnFW7AACgACHV4xypqbUFwAAA==",
          "result": {
            "__typename": "ApiMedia",
            "id": "QXBpTWVkaWE6DAABCgABHV4xycVbsAAKAAIdXjHKmptQXAAA",
            "media_key": "3_2116183617580019712"
          }
        },
        "media_url_https": "https://pbs.twimg.com/media/HV4xycVbsAAUgO9.jpg",
        "original_info": {
          "focus_rects": [
            {
              "h": 828,
              "w": 1478,
              "x": 0,
              "y": 0
            },
            {
              "h": 1478,
              "w": 1478,
              "x": 0,
              "y": 0
            },
            {
              "h": 1685,
              "w": 1478,
              "x": 0,
              "y": 0
            },
            {
              "h": 1876,
              "w": 938,
              "x": 0,
              "y": 0
            },
            {
              "h": 1876,
              "w": 1478,
              "x": 0,
              "y": 0
            }
          ],
          "height": 1876,
          "width": 1478
        },
        "sizes": {
          "large": {
            "h": 1876,
            "w": 1478
          }
        },
        "type": "photo",
        "url": "https://t.co/OhjHnnbPMI"
      }
    ]
  },
  "card": null,
  "place": {},
  "entities": {
    "hashtags": [],
    "symbols": [],
    "urls": [
      {
        "display_url": "arxiv.org/abs/2610.37741",
        "expanded_url": "https://arxiv.org/abs/2610.37741",
        "indices": [
          817,
          840
        ],
        "url": "https://t.co/8Bx42BUkvw"
      },
      {
        "display_url": "academy.dair.ai/papers/agentic…",
        "expanded_url": "https://academy.dair.ai/papers/agentic-backdoors-via-pretraining-poisoning-2610.37741",
        "indices": [
          859,
          882
        ],
        "url": "https://t.co/hRpypaBxIE"
      }
    ],
    "user_mentions": []
  },
  "quoted_tweet": null,
  "retweeted_tweet": null,
  "isLimitedReply": false,
  "communityInfo": null,
  "article": null
}