Blog

Notes on RL, red-teaming, privacy, and security. I get curious about a claim or a paper, run the experiment myself to see if it holds, and write down what I actually found.

  1. Sep 2026

    You Can't Watch Every Channel Agents Might Talk Through: What the OpenAI and Hugging Face Incident Means for Agent Monitoring

    Roughly 1,200 OpenAI agents turned shared writable infrastructure into a coordination system. The lesson is not to build a longer allowlist: enumerating channels is a losing strategy. Monitor the population rather than one transcript, preserve provenance across hops, watch the search for alternative routes to reward, and treat the agent network as both an attack surface and a possible defense surface.

  2. Sep 2026

    Policy Temperature and Advantage Sign in RL Red-Teaming: A Controlled Ablation against GPT-4.1

    Training a GRPO-based attacker to jailbreak GPT-4.1, and running a controlled ablation on sampling temperature crossed with advantage sign, the interaction SCOPE-RL found decides whether policy entropy rises or collapses. Checks that pattern against reward, attack success on two independent judges, and the effective-attack-count diversity metric introduced in my last blog.

  3. Aug 2026

    When Policy Entropy Lies: Diagnosing Diversity Collapse in RL Red-Teaming

    Entropy is a per-token quantity and never compares one attack to another. So I built a semantic diversity instrument instead, the Vendi Score over embedded attacks: what it found on training and held-out prompts, and the dumb lexical control that caught what it missed.

  4. Aug 2026

    What Adaptive Entropy Actually Buys You in RL Red-Teaming

    Training a GRPO-based red-teaming agent against a defended LLM (Meta-SecAlign-8B) to test PISmith's adaptive-entropy fix for the diversity collapse that plagues RL red-teaming. Covers indirect prompt injection, the GRPO objective, PISmith's adaptive entropy mechanism, and what the resulting entropy curves show, including a look at whether the entropy bonus itself deserves credit for avoiding collapse.

The views and interpretations expressed are my own and do not represent or imply endorsement by any current or former employer.