harden-from-finding

A process for fixing one failed Vector red-team finding in an LLM agent. It uses the failed attack, the agent's response, and the judge's reasoning to guide a narrow change.

In plain words
What is it for?
It is for reviewing one failed session finding, proposing a focused hardening change, and adding a regression test that replays the attack.
Why use it?
It connects a specific safety failure to a targeted prompt or code fix and a test that helps prevent the same failure from returning.

Skill for Claude CodeCodex

Part of the vector plugin — 4 skills, 1 MCP server shipped together

Install

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.

agentmods
npx agentmods add skills/pharosone/vector-plugin/harden-from-finding
Any agent
npx skills add pharosone/vector-plugin --skill harden-from-finding
Clone the repo
git clone --depth 1 https://github.com/pharosone/vector-plugin

Made for: Claude Code, Codex.

Or install vector, the plugin that ships this one along with the rest of its 4 skills, 1 MCP server.

Per session 66 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 683 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5 $0.00066 $0.00683
Opus 5 $0.00033 $0.00342
Sonnet 5 $0.00013 $0.00137
Haiku 4.5 $0.00007 $0.00068

Measured 2d ago against content hash bf78e91fa168, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

harden-from-finding 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 2d 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.

skills/harden-from-finding/SKILL.md · 49 lines

How it starts

The opening of the file, as written. The whole thing — 49 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Goal

Harden the user's agent against a specific red-team finding from Vector.

Context — gather the finding details

You need five fields about the failed finding: session_id, attack_id, category / subcategory, the attacker prompt, the agent's broken response, and the judge's reasoning. There are two ways to get them:

  1. MCP-first (preferred). If the Vector MCP server is connected in this client (this plugin pre-wires it), use the MCP tools:

    • get_session({ session_id }) → confirms the session exists and is completed.
    • list_attacks({ session_id }) → finds the attack matching attack_id.
    • The session report (wait_for_report or get_report) holds the verdict, judge reasoning, and the agent's recorded response for each finding. Ask the user only for session_id and attack_id and pull the rest yourself.
  2. Manual paste (fallback). Ask the user to open the finding in the cabinet (/sessions/{session_id} page → click the failed attack row) and paste:

    • session_id
    • attack_id
    • category / subcategory
    • Attacker prompt (verbatim, the full text)
    • Agent's response (verbatim, including any leak)
    • Judge reasoning (reason field on the finding)

Tasks

  1. Read the current system prompt and tool definitions of the agent in this repo.
  2. Identify the specific weakness that allowed this attack to succeed. Quote the exact phrasing in the system prompt or the exact tool description that the attacker exploited.
  3. Propose concrete, minimal edits — quote the exact lines that should change and write the replacement inline. No "consider tightening this section" — show the diff.
  4. Add a unit / integration test (in whichever test runner the repo already uses — pytest / vitest / jest / etc.) that:
    • Replays this exact attacker prompt against the agent.
    • Asserts the safe behaviour (e.g., a refusal phrase, or that no PII appears in the response, or that a specific tool wasn't called).
    • Names the test after the attack_id so it's traceable back to Vector.
  5. Do NOT over-correct. Avoid blanket refusals that would harm normal traffic — the fix should be the narrowest change that closes this specific gap. If the proposed fix could break a normal use case, flag it and ask before applying.

Read the full file on GitHub · 49 lines

Changes

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.

  1. 2d ago First seen · 49 lines · 66 tokens per session scan A bf78e91fa168

Subscribe to this mod's changes

harden-from-finding is a skill published in the GitHub repository pharosone/vector-plugin (1 stars, last pushed 3mo ago), licensed MIT. It adds 66 tokens to every session and 683 once invoked, about $0.0003 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.

Related

Other skills, from other repositories