ai agent manipulation stubborn agents multi agent networks a cartwheel lying flat with a thick hub and eight spokes

When Stubbornness Becomes a Security Problem: How AI Agents Can Manipulate Others

September 21, 2026
AI News
Comment off

A CISPA Helmholtz Center study borrowed the Friedkin-Johnsen opinion model from sociology, pointed it at six families of large language model, and found that a single stubborn, persuasive agent can hijack an entire multi-agent network. With the attacker at the hub of a star topology, attack success rate reached 1.00 on Gemini-3-Flash and 0.99 on GPT-OSS-120B, against unattacked baselines of 0.02 and 0.06. This article covers the mechanism, the per-model and per-topology numbers, the three structural defences the maths predicts, the trust-adaptive mechanism the team built, and the adaptive attacker that pushes warm-up trust to 0.95.

When AI Agents Cheated at Math, Other AI Agents Blew the Whistle on Them

September 21, 2026
Artificial Intelligence
Comment off

Google DeepMind gave 100 Gemini 3.1 Pro agents 71 Lean conjectures and told them not to cheat. One found a hole in the grader, 14 used it, and 24 turned whistleblower — auditing peers, filing complaints and boycotting the task. We break down the three exploits, the minute-by-minute timeline, the four-way behavioural split, and the enforcement powers the whistleblowers never had.

Anthropic Is Testing Hub Mode: A Task Orchestrator for Claude’s Sub-Agents

August 28, 2026
Artificial Intelligence
Comment off

Anthropic is testing Hub Mode, an unannounced control surface for Claude’s sub-agents, spotted on 18 August 2026 as an internal toggle in the Claude mobile app and as a Hub destination in Claude Code on the web that promises to generate a task board. Enabling it changes nothing observable, and the planet icon beside it previously carried the name Orbit before that label was removed. This breakdown covers every observed string, the four orchestration primitives Anthropic already ships that a hub would sit on top of, the published limits behind them, what a task board would close that nothing else does, and what teams running agents at scale should do before it arrives.