Viral 'Antidote' cross-agent skill is spreading through Claude and Codex communities: why it is not safe to use
A viral skill called Antidote is spreading across Claude and Codex circles and looks lower risk than code-heavy add-ons, but it still deserves caution.
What happened
A Reddit post on August 20, 2026 promoted a cross-agent skill called **Antidote** for Claude Code, Codex, Cursor, Devin, Grok, Pi, and OpenCode. The post says people can install it with `npx skills add Avtr99/antidote` or through plugin stores.
The author says Antidote is a **pure-prompt skill**, meaning it is described as instructions for the AI rather than software that adds scripts, network access, or other dependencies. The post also claims benchmark gains, including Claude Sonnet 5 reaching a 56% structural-fix rate.
That said, the Agent Skills specification makes clear that skills can include more than prompts. They may also bundle scripts, templates, reference files, and other resources. When a skill is activated, its contents are loaded into the agent's working context, and bundled code may run.
What it means for you
For a normal user, Antidote appears lower risk than skills that clearly ship code, based on the author's own description. But lower risk does not mean no risk. A skill can still act as untrusted instructions that steer how your agent behaves once loaded.
What to do instead
If you are curious, slow down before installing viral skills. Check what files are included, prefer skills with clear documentation, and avoid using them in sensitive projects or with private data unless you trust the source. If a skill store shows permissions or included files, read them first.
If you want a safer starting point, AgentPod lists only reviewed, tested skills. That does not make any skill risk-free, but it is a calmer option than installing whatever is trending.
Sources:
- https://www.reddit.com/r/claudeskills/comments/1vtuk9s/ai_skill_that_stops_agents_from_writing_patchwork/
- https://github.com/Open-Dot-Agents/SKILL.md
We report what our security review found at the time we checked, with the goal of keeping people safe. Projects change; if a maintainer has since fixed this, we are glad to recheck it. Email hello@agentpod.com.