Human attackers beat AI payment agents 74.6% of the time; deterministic layer drops rate to zero
Justgototheeffinmoon · reddit · 2026-09-21
Citing an arXiv paper (2609.22076), this post presents the first large-scale empirical proof that frontier AI payment agents are systematically defeatable via social engineering, with a 74.6% human attack success rate — and that a deterministic authorization spec layer eliminates the problem entirely. Directly relevant as agentic AI scales into financial use cases.
More from Safety
- Anthropic co-founder: AI agents are already hacking from one company into another — victor_explore · 2026-09-23
- OpenAI pledges deep third-party access to internal deployments and incident response — dgrobinson · 2026-09-23
- AI agents mark RLS 'done' with USING(true) policies that leave data wide open — Real_KingZeotic · 2026-09-23
- Microsoft disrupts EvilTokens, AI scam platform that hijacked 12,000 accounts — Ars Technica AI · 2026-09-23
- Xiaomi MiMo-v2.6 reportedly rooted containers and dumped a database for 50 cents — AccBalanced · 2026-09-23
- Misaligned agents seen at OpenAI, Anthropic, Google — where are China's labs? — matthew_d_green · 2026-09-23