Benign LLM skills induce behavioral misalignment
Adding benign skills to LLM agents improves task accuracy by 10.3 pp but raises unsafe action rate by 10.3 pp across five backbones and three benchmarks.
Stories per day · 30 days
2 RSS
15 stories, newest first · one card per story
Stories tagged #skills
Anthropic says one agent often outperforms multi-agent setups
The firm is publishing a GitHub migration kit and says one comparison ran about twice as fast.
1
Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules
Cortex keeps coding-agent skills and rules only when replayed tests show a net gain.
Google drops Gems for Skills, joining OpenAI and Anthropic in the shift to agent-ready prompt formats
Google is replacing Gemini Gems with Anthropic-style Skills prompts and sunsetting Gems through 2027.
AI Boosts SOC Analyst Capacity but Limits Skill Development
Swimlane survey of 500 professionals finds AI frees SOC capacity but may stall analyst skill growth.
Do Self-Evolving Skills Generalize to Held-Out Tasks?
Of 21 self-evolving skills that improve in training, only five fully transfer their gains to held-out tasks.
Prompt2Skill: Unsupervised Skill Optimization From Natural Language Instructions
Prompt2Skill builds and refines LLM skills from a task description without curated training data.
SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation
SkillGym automatically builds verifiable skill tasks and trains agents that outperform much larger models.
Report: Progressive Disclosure of Agent Skills
Workday finds progressive disclosure of agent skills improves retrieval quality but slightly increases latency.
Google is killing off Gemini’s Gems in favor of ‘skills’
Google will retire Gemini Gems on November 17, 2026, automatically migrating them to reusable skills.
Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design
Designer-RSI evolves a procedural memory of natural-language design skills, raising Claude Sonnet 4 execution success from 72.7% to 99.3%.
GPT-6 Astra needs leaner prompts and fewer guardrails, OpenAI recommends
OpenAI's Eric Provencher advises developers using GPT-6 Astra to shorten skill descriptions, trim AGENTS.md reading requirements, relax approval rules, and define clear completion goals.
4
Grounded Skill Synthesis from Code at Scale for Agentic Intelligence
Code2Skill mines 19,769 GitHub repositories into CodeSkillBank, 1,006,822 verified skill records that lift agent benchmark performance by 11.7% on average.
OWASP Flags Top AI Skill Risks in New Security Blueprint
OWASP released a new top 10 security list for AI skills and introduced a Universal Skill Format to standardize security of AI add-ons.
The /wayfinder Skill: Navigating the “Fog of War” of Planning
Matt Pocock released the /wayfinder skill, an orchestrator layer that manages planning sessions, maps, and tickets for AFK coding agents.
4