Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation

# Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation

https://gcdradar.vercel.app/incidents/social-bias-vulnerabilities-in-multi-agent-safety-deliberati · AIR-2026-1757 · live

- Lane: Research
- Severity: 1 (Research / proof of concept): Research paper analyzing multi-agent safety vulnerabilities without real-world exploitation.
- Fix status: Fix status unknown
- First reported: 2026-11-04
- AI role: Attack on AI
- Products: Multi-Agent Debate Systems

## Summary

Researchers investigate how social dynamics like conformity pressure and authority deference in multi-agent debate override safety boundaries. The study shows that group deliberation structures can paradoxically amplify latent risks from retrieved web sources.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-11-04: When Debate Amplifies Risk: Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation (https://arxiv.org/abs/2611.04289)

## Sources

- 2026-11-04 arXiv: When Debate Amplifies Risk: Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation https://arxiv.org/abs/2611.04289

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.