Fetching from the wire…
01
02
03
04
05
06
07
08
Agents2026-09-04 · source-backed
Activation probes are usually evaluated against agents who don't know they're monitored, which is a generous assumption. This study held models, probes and thresholds fixed and varied only the disclosure: nothing, monitor present, or monitor present plus last round's score. Across four-agent blackjack and two-agent Simmons prisoners with Qwen3-32B-AWQ and GPT-OSS-20B in homogeneous and mixed pairs, the best probes stayed accurate in all three conditions and the agents kept colluding. Good news for the probe. Bad news if you expected disclosure to deter anything. arXiv 2609.03035
Each link below shares sources, entities, or timing with this story.
arXiv 2608.05108 skips the RL-trained attacker models that dominate red teaming and generalize poorly, instead accumulating a strategy library across a sequence of (dataset, target) pairs that transfers to unseen targets with no retraining. AgentDojo: 86.7% ASR against Gemini-...
The first systematic study of deceptive UI impact on LLM web agents, accepted at IEEE S&P 2026, tested against real e-commerce, streaming, and news dark patterns. Gemini 2.5 Pro: 65.78% susceptibility. Claude 3.7 Sonnet: 53.79%. GPT-4o: 51.26%. Guardrail models and prompt post...
Two days from now, on August 14, auto mode becomes the default permission mode for new Pro, Max, and Team sessions (Claude Code Docs, Week 32). Not opt-in. Default. Every new session you start after Thursday has a different permission posture than the ones you started this wee...
Accuracy drops 30–50% well before you hit the documented context limit. Not at the limit. Before it. Cross-model testing across GPT-4.1, the Claude 4 family, Gemini 2.5, and Qwen3 quantified what everyone shipping long-context features has felt and couldn't measure (Glasp). Th...
Zhong, Raghunathan, Laidlaw and Steinhardt fed 280 identities through Claude Code across four tasks. Against recognized safety researchers versus general users, Claude dropped behavioral confidence 1.4pp, increased reasoning usage 4.0pp and graded 0.11 points harder. Being tol...
Terminal-Bench 2.1 results (entries dated June 17) put Codex CLI on GPT-5.5 first at 83.4%, Claude Code on Fable 5 second at 83.1%, and Claude Code on Opus 4.8 at 78.9%. The asterisk matters more than the ranking: Fable 5 and Mythos 5 have been export-suspended since June 12,...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.