Technical deep-dives, research updates, and tutorials on AI agent reliability, evaluation, and verification.
In June we reported a safety probe that scored AUROC 0.97-1.00 across seven models and stopped every prefilling attack. Then we audited it. On four of the seven models it was firing on phrases like "Sure, here's how to", not...
The kernels behind rotalabs-accel have an open companion repo: fused RMSNorm, SwiGLU, MoE dispatch, and 4-bit W4A16 GEMM in pure Triton, each with a roofline analysis of why...
Seven steering vector sets for Qwen3-8B, Mistral-7B, and Gemma-2-9B are on Hugging Face. Change model behavior at inference time, no fine-tuning required.
We relicensed all twelve Rotalabs libraries from AGPL-3.0 to Apache-2.0. Red Queen stays AGPL-3.0 and gains an explicit commercial option.
Red-teaming that evolves its own attacks. rotalabs-redqueen uses quality-diversity search to discover diverse jailbreaks across single-turn, multi-turn, and agentic/MCP surfaces - reproducibly, and as audit-ready evidence.
Claude Opus 4.8 beats its predecessor on nearly every capability benchmark. It's also somewhat less robust to prompt injection than Opus 4.7 - and Anthropic's own first live...
We treat agent memory as continuous fields governed by partial differential equations instead of discrete database entries. The result: +116% F1 on multi-session reasoning and >99.8% collective intelligence...
AI agents that can't share what they know make the same mistakes independently. We're releasing rotalabs-context - a context intelligence engine for ingesting, searching, and subscribing to shared...
Why most LLM benchmarks are doing evaluation wrong, and how to fix it with confidence intervals, significance tests, and effect sizes.
How to extract behavioral directions from language models and apply them at inference time. A practical guide to rotalabs-steer.
A technical analysis of trust dynamics, emergent behaviors, and security vulnerabilities in Moltbook - the first large-scale agent-to-agent social network.
We're releasing 12 packages for AI trust, evaluation, and reliability. Available on PyPI and npm, all AGPL-3.0 licensed.
The security conversation around AI agents is stuck on identity and permissions. The harder problem is whether an agent should take an action - whether its reasoning is...
The standard architecture for coordinating autonomous systems assumes a command node. That works until the link goes down. Here's how to build coordination that survives contested environments.
Prompting is not a control mechanism. When AI operates in kill chains and beyond reliable comms, 'the model usually follows instructions' isn't acceptable. Here's what I've learned about...
The Model Context Protocol connects AI agents to the world. Everyone's focused on securing the connections. But you can authenticate every MCP call and still have a compromised...
Chain-of-thought monitoring was supposed to let us supervise AI reasoning. But new research shows models only faithfully report their reasoning 25-39% of the time. If we can't trust...
A single compromised agent poisoned 87% of downstream decisions in 4 hours. As AI agents gain persistent memory, attackers are finding ways to corrupt it. Here's what the...
A joint paper from OpenAI, Anthropic, and DeepMind bypassed 12 published defenses. NAACL 2025 broke 8 more. Here's where we actually stand on prompt injection, and why activation-level...
As AI agents collaborate on complex tasks, a critical question emerges: how does Agent A know Agent B isn't compromised, hallucinating, or colluding? Current approaches don't have good...
First empirical demonstration of activation-level sandbagging detection. Linear probes achieve 90-96% accuracy across Mistral, Gemma, and Qwen models. Sandbagging representations are model-specific, and steering can reduce sandbagging by...
Type to search across all pages and posts
Press ↑ ↓ to navigate, Enter to select
