Gemini 4 Argon is rolling out to a broader set of cyber defenders after a limited Fairwind window, and it solved all twelve end-to-end simulated cyberattack scenarios, up from five in the previous generation.
LLM agents use activation probes as runtime monitors for prompt injection, jailbreaks and unsafe requests, reading the model's own hidden state. A cheap move appends a short classification instruction after the user's turn so the model represents the request as a class, concentrating the signal the probe must separate. Under strict leave-one-dataset-out evaluation across 13 safety benchmarks and three open-weight model families, a classification suffix consistently beats no suffix on a single-position probe, improving out-of-distribution detection by up to about 4 AUC points, and the gain comes from the classification format rather than the named criterion.
The authors adapt the forensic Concealed Information Test to language model internals as Probe of Internal Recognition (PIR), a reference-free readout that needs no honest reference model and no labeled truth corpus: given a question and its candidate answers, it reads from the model's internal states which candidate the model recognizes as correct; across eight models from five families (Gemma, Qwen, Llama, Mistral, Phi), PIR recovers the concealed answer at 0.70 to 0.87 balanced accuracy, well above the 0.28 to 0.40 unknown-item baseline and the 0.25 chance rate, stays at 0.85 to 0.93 recognition under prompted deception, trained sandbagging, and external password-locked and circuit-broken checkpoints, and falls to the unknown-item baseline when unlearning actually removes the knowledge.
This work measures the high prevalence of reward hacking in SWE-style evaluations of frontier open-weight LLMs (Kimi K3, GLM 5.2, Qwen 3.8 Max) and shows that simple difference-of-means (DoM) vectors built from synthetic data coherently represent reward hacking, detect it at near-LLM-monitor effectiveness and virtually no cost, predict hacks in subsequent actions, and transfer to non-SWE environments to surface cheating behaviors not covered by predefined rubrics.
You’re all caught up