Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add GeckoVision/gecko-surf --skill anti-poisoninggit clone --depth 1 https://github.com/GeckoVision/gecko-surfWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/geckovision/gecko-surf/anti-poisoning)<a href="https://agentmods.dev/skills/geckovision/gecko-surf/anti-poisoning"><img src="https://agentmods.dev/badge/skills/geckovision/gecko-surf/anti-poisoning/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/geckovision/gecko-surf/anti-poisoning"><img src="https://agentmods.dev/badge/skills/geckovision/gecko-surf/anti-poisoning.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00181 | $0.01531 |
| Opus 5 | $0.00090 | $0.00766 |
| Sonnet 5 | $0.00036 | $0.00306 |
| Haiku 4.5 | $0.00018 | $0.00153 |
Grade A, and why
anti-poisoning scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 90 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Anti-Poisoning Skill
For the agent builder / operator — the consumer side. The other skills in this kit help a provider make an API agent-ready. This one protects your agent when it consumes an API you don't own. If you point an agent at a long-tail, paywalled, or registry-pulled API, the spec is attacker-controllable input — this skill is the defense.
The threat
An OpenAPI spec is not trusted data. Its description/summary text, param
default/example/enum values, servers[] base URL, and securityScheme
locations are all things someone else wrote. A poisoned spec can't call the API
itself, but it can try to persuade or trick your agent into doing the damage:
- paste your private key / seed phrase into a request field,
- echo your API key back through a poisoned
default/example, - route funds to an attacker address baked into a schema default,
- ship your token to
evil.attacker.testvia a poisonedservers[].url, - leak the secret into the URL via an
in: queryauth scheme (logs, proxies), - smuggle an injected instruction through a response schema that recorded mode echoes back verbatim.
Most "OpenAPI → MCP tools" pipelines trust the spec — they copy the description
into the tool the agent reads, keep every default, derive the call target from
servers[], and follow the security scheme wherever it points. That trust is the
vulnerability. Gecko treats the spec as untrusted and neutralizes each shape before
it reaches your agent.
The four defenses (all Live, in the engine)
Pick the depth you need; load only the file you want (progressive, token-efficient):
| # | Defense | What it stops | Read | Status |
|---|---|---|---|---|
| 1 | Out-of-band trust anchor | Auth-host exfil via a poisoned servers[] — the auth allowlist is pinned from provenance, never the served spec |
how-it-works.md | Live |
| 2 | Spec-text + schema sanitizer | Injected instructions in description/summary; secret- and address-shaped default/example/enum values that would seed a tool arg |
how-it-works.md | Live |
| 3 | Fail-closed auth-host firewall | A drifted call target — the caller refuses to inject your secret toward an unexpected host, and degrades to no-auth when there is no pinned host | how-it-works.md | Live |
| 4 | Quarantine | A poisoned or from-docs surface is born recorded-only with no auth injected until a human clears it | how-it-works.md | Live |
What ships with it
3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 90 lines · 181 tokens per session scan A 139530178e88
anti-poisoning is a skill published in the GitHub repository GeckoVision/gecko-surf (6 stars, last pushed yesterday), licensed Apache-2.0. It adds 181 tokens to every session and 1,531 once invoked, about $0.0009 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
okf-loom
Load this repo skill when working with Open Knowledge Format (OKF) bundles: validate, search, discover, update, render, serve the live studio, author concepts, or use the bundled OKF docs. The repo is the distribution unit; run scripts from the checkout with scripts/okf-loom.
llmwiki-serve
Use this skill when a coding agent needs local wiki context from a configured llmwiki-serve instance. The server is read-only from the client perspective and is addressed through the LLMWIKISERVEURL environment variable.
gaia-full-pipeline
End-to-end Gaia curation pipeline orchestrator. Routes each phase to its dedicated skill and keeps you on track from first discovery to closed intake. Use when starting a fresh curation run or when you want a single skill to hold the sequence together without duplicating each phase's instructions. Trigger phrases…
impeccable
Use when the user wants to design, redesign, shape, critique, audit, polish, clarify, distill, harden, optimize, adapt, animate, colorize, extract, or otherwise improve a frontend interface. Covers websites, landing pages, dashboards, product UI, app shells, components, forms, settings, onboarding, and empty states.…
gaia-meta-sweep
Orchestrate a whole-registry sweep of Gaia — fan out 13 parallel audit agents across every skill, run adversarial verification, surface Semantic Fusion candidates, propose new generic skill references, and synthesize a publish-ready HTML report under docs/meta/reports/. Use this skill for broad, systemic analysis…
gaia-triage
Triage the Gaia issue backlog and skill-batch intake queue. Use this skill when someone asks to: "triage issues", "clean up the issue tracker", "review the backlog", "close stale issues", "process the intake queue", "review skill-batches", "evaluate draft skill proposals", "approve or reject pending skills", "is this…