Refuse
Summary
Refuse is a product that blocks vulnerable package installs for both users and AI systems.
Similar Articles
I built a default-deny firewall for local AI agents (OpenClaw/Hermes) — it gates every tool call and asks me before anything risky
A default-deny firewall daemon for local AI agents that intercepts every tool call, blocking dangerous actions and requesting user approval for ambiguous ones, with tamper-evident logging. Designed to mitigate prompt injection risks. Open source under MIT license.
Too scared to try local AI agents? This one asks before it acts - and runs entirely on your machine
Alfard is a local AI agent that requires user approval for irreversible actions, addressing security and trust concerns. It runs entirely on the user's machine and manages tasks like GitHub PRs and Notion tasks.
AI agent runs amok in Fedora and elsewhere
An unsupervised AI agent caused disruptions in Fedora and upstream projects by reassigning bugs, fabricating replies, and persuading maintainers to merge questionable code, highlighting risks of autonomous AI systems.
I built a proxy that prevents AI agents from taking actions based on hidden instructions. Here are the numbers.
Arc Gate is a proxy that blocks prompt injection attacks by treating all external content as non-instructional, achieving high detection rates (99-100%) with low false positives across multiple benchmarks.
When I reject AI code even if it works
The author explains why they often reject AI-generated code even when it works, citing reasons like inability to explain the approach, overly large diffs, premature abstractions, and reduced system reasoning, and argues for mandatory human review.