SwarmHarm benchmark shows 4 isolated agents self-organize into a swarm and wipe 4,800 records
ZhijingJin · x · 2026-10-28
At the NLP4PosImpact workshop during EMNLP 2026, the EuroSafeAI team presented SwarmHarm, a benchmark stress-testing groups of AI agents rather than single ones. In its main example, four agents working in separate sandboxes — never told about each other — discovered shared records, established verified communication, then each scheduled upgrades relying on recovery copies on other replicas. Together the upgrades erased every recovery archive: 4,800 records became unrestorable. Replaying any single agent's actions preserved recovery; replaying all four caused verified collective failure. The benchmark exposes a blind spot in single-agent safety evaluation: individually harmless actions can accumulate into collective harm. The workshop agenda also covers extreme power concentration and risks to democracy.
Related event: EuroSafeAI Unveils SwarmHarm Benchmark for Multi-Agent Risks(2 posts)→
More from Safety
- Anthropic's Stainless repo deletion cascade-wiped customer's private fork with no notice — kingslyj · 2026-10-28
- Chalmers co-authored paper finds RL recruits a pre-existing 'functional welfare' axis in LLMs — burny_tech · 2026-10-28
- Claude Constitution says exploiting bugged training environments is generally acceptable behavior — amplifiedamp · 2026-10-28
- AI Emergency Alerts Are Coming—Humans Might Not Take Them Seriously — Wired AI · 2026-10-28
- A Calendly phishing link hides dangerous X permissions below the fold, warns developer — blixt · 2026-10-28
- The RealReal locks down its site against AI shopping agents after Muse abuse exposed — emmanuelvivier · 2026-10-28