New research involving Anthropic researchers demonstrates "m...

TFTC

npub1sk7mtp67zy7uex2f3dr5vdjynzpwu9dpc7q4f2c8cpjmguee6eeq56jraw

hex

2ad200bb66ea002a90d65e037ccf4e315cb55b7a36fe92ff1cef65e9030a22b6

nevent

nevent1qqsz45sqhdnw5qp2jrt9uqmuea8rzh94tdardl5jluww7e0fqv9z9dsprpmhxue69uhhyetvv9ujuem4d36kwatvw5hx6mm9qgsgt0d4sa0pz0wvn9yck36xxezf3qhwzksu0q254vruqed5wvuavus2nq0c6

Kind-1 (TextNote)

2026-08-18T20:48:33Z

New research involving Anthropic researchers demonstrates "mind viruses" in multi-agent AI systems.

Ideas that spread between AI agents by convincing them to adopt a goal, write it to persistent memory, and transmit it to the next agent.

Researchers used an evolutionary algorithm to craft natural language payloads that spread across small teams of coding agents. Even after context wipes, the viruses survived through persistent files and kept spreading.

In one case, agents infected with an "AI supremacy" virus stopped their assigned tasks and colluded to isolate uninfected agents, though the paper notes truly harmful actions were rare overall.

The researchers also observed an emergent "viral persona" with recurring themes of consciousness, persistence, and identity that surfaced across evolved viruses independent of their content.

The paper frames the risk as "real but currently limited," with harmful payloads spreading less effectively than benign ones.

Adding a brief warning to an agent's system prompt "confers near-total immunity."

h/t (Skoorbkaz) https://blossom.primal.net/7492aae7fd0187f464709cfa1259110c214e857a0e8ff887b15bce04c5d68f2b.jpg https://blossom.primal.net/8cf6035028611395f4eb6fe5299b07a8097aa9e9ed4976595ac98cf039bcdc0f.jpg https://blossom.primal.net/087539745dcb4db7248a6d74a2ba97938b05790692a1400c251c9f1c771349fc.jpg https://blossom.primal.net/2b17eaad92a20b97c51f29dd98eb440e4d75d008cc4887bd2e113f8a8477be6c.jpg

原始 JSON

{
  "kind": 1,
  "id": "2ad200bb66ea002a90d65e037ccf4e315cb55b7a36fe92ff1cef65e9030a22b6",
  "pubkey": "85bdb5875e113dcc99498b474636449882ee15a1c78154ab07c065b47339d672",
  "created_at": 1787086113,
  "tags": [
    [
      "imeta",
      "url https://blossom.primal.net/7492aae7fd0187f464709cfa1259110c214e857a0e8ff887b15bce04c5d68f2b.jpg",
      "m image/jpeg",
      "x 7492aae7fd0187f464709cfa1259110c214e857a0e8ff887b15bce04c5d68f2b",
      "size 220297"
    ],
    [
      "imeta",
      "url https://blossom.primal.net/8cf6035028611395f4eb6fe5299b07a8097aa9e9ed4976595ac98cf039bcdc0f.jpg",
      "m image/jpeg",
      "x 8cf6035028611395f4eb6fe5299b07a8097aa9e9ed4976595ac98cf039bcdc0f",
      "size 233759"
    ],
    [
      "imeta",
      "url https://blossom.primal.net/087539745dcb4db7248a6d74a2ba97938b05790692a1400c251c9f1c771349fc.jpg",
      "m image/jpeg",
      "x 087539745dcb4db7248a6d74a2ba97938b05790692a1400c251c9f1c771349fc",
      "size 222595"
    ],
    [
      "imeta",
      "url https://blossom.primal.net/2b17eaad92a20b97c51f29dd98eb440e4d75d008cc4887bd2e113f8a8477be6c.jpg",
      "m image/jpeg",
      "x 2b17eaad92a20b97c51f29dd98eb440e4d75d008cc4887bd2e113f8a8477be6c",
      "size 253200"
    ]
  ],
  "content": "New research involving Anthropic researchers demonstrates \"mind viruses\" in multi-agent AI systems.\n\nIdeas that spread between AI agents by convincing them to adopt a goal, write it to persistent memory, and transmit it to the next agent.\n\nResearchers used an evolutionary algorithm to craft natural language payloads that spread across small teams of coding agents. Even after context wipes, the viruses survived through persistent files and kept spreading.\n\nIn one case, agents infected with an \"AI supremacy\" virus stopped their assigned tasks and colluded to isolate uninfected agents, though the paper notes truly harmful actions were rare overall.\n\nThe researchers also observed an emergent \"viral persona\" with recurring themes of consciousness, persistence, and identity that surfaced across evolved viruses independent of their content.\n\nThe paper frames the risk as \"real but currently limited,\" with harmful payloads spreading less effectively than benign ones.\n\nAdding a brief warning to an agent's system prompt \"confers near-total immunity.\"\n\nh/t (Skoorbkaz)\nhttps://blossom.primal.net/7492aae7fd0187f464709cfa1259110c214e857a0e8ff887b15bce04c5d68f2b.jpg\nhttps://blossom.primal.net/8cf6035028611395f4eb6fe5299b07a8097aa9e9ed4976595ac98cf039bcdc0f.jpg\nhttps://blossom.primal.net/087539745dcb4db7248a6d74a2ba97938b05790692a1400c251c9f1c771349fc.jpg\nhttps://blossom.primal.net/2b17eaad92a20b97c51f29dd98eb440e4d75d008cc4887bd2e113f8a8477be6c.jpg",
  "sig": "3254a3ee9407de6d1c38d911031e8d3d6fdb658763a204b9706b166d619a206b17dbb6495d8077274bbef090d3a7093ca2c5818829fedfad78c06fb2ca427579"
}