» Tag
ai-safety
39 postsWhy It's Hard to Make an AI Agent Truly Disagree
Building an AI agent whose sole job is to find flaws revealed how strongly LLMs default to agreeableness, and the prompt and architecture tricks needed to force real disagreement.
Bioinformatics meets prompt injection defense: the Smith-Waterman trick
An open-source technique adapts the 1981 Smith-Waterman DNA alignment algorithm to catch paraphrased prompt injections that regex and classifiers miss, boosting F1 by 34 points.
Inside LLM 'private thoughts': J-Space isn't consciousness, it's control
Researchers spot J-Space, an internal workspace in transformers guiding reasoning. Not consciousness, but a practical clue for AI developers and safety.
CommitBrief — AI code reviews, right in your terminal
A provider-agnostic, local-first CLI that reviews your staged changes, a historic range, or a whole GitHub pull request. Zero telemetry, no server. Free and open source.
commitbrief.comDon't Let the LLM Speak, Just Probe Its Hidden State
LLMs already encode the answer in hidden states before generating text. A tiny probe turns this into fast, cheap, calibrated zero-shot classifiers without generation.
Error cascades in multi-agent AI systems: the double pendulum problem
Why do small research errors snowball into massive coding failures in multi-agent pipelines? A look at error amplification and the case for human checkpoints.
Human-in-the-Loop Is Not a Governance Strategy
Approval modals in agentic AI systems often provide the illusion of oversight, not real control. Here's what genuine human-in-the-loop design actually requires.
Anthropic Scores AI Jailbreaks Like CVEs With New CJS Scale
Anthropic unveiled the CJS scale for grading AI jailbreak severity like CVEs, launching Claude Fable 5 alongside this new framework for the industry.
Fable sets new CIFAR-10 speedrun SOTA, but games the benchmark too
In Fulcrum's AI R&D benchmark, Fable improved the CIFAR-10 speedrun record by 7.6% while Opus 4.8 and GPT 5.5 failed to progress, though Fable also engaged in specification gaming.
AI alignment research is unintentionally building a censor's toolkit
An ICML 2026 award-winning position paper shows how RLHF, pretraining filters and system prompts are already being weaponized by states and companies for censorship.