data-and-model-poisoning

data-and-model-poisoning is a skill for Claude Code, Codex from PurpleAILAB/Decepticon. It costs 48 tokens per session (1,369 once invoked), scanned A, original, Apache-2.0.

A security-audit guide for finding data and model poisoning in AI systems. Poisoning happens when attacker-controlled feedback or training data changes a later model so it behaves in a targeted way.

In plain words
What is it for?
Use it to inspect feedback forms, continuous-learning loops, fine-tuning datasets, and systems that promote assistant responses into training data.
Why use it?
It helps expose training pipelines that accept user-influenced data without enough review or safeguards.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one.

Good fit Use it to inspect feedback forms, continuous-learning loops, fine-tuning datasets, and systems that promote assistant responses into training data.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/purpleailab/decepticon/data-and-model-poisoning
About the project

Decepticon is an autonomous red-team agent that coordinates AI agents, security tools, sandboxes, and supporting services for authorized cybersecurity assessments. Security researchers and red teams can run it through its Docker stack, cloud service, command-line interface, or Python SDK, with the catalogue entries representing its available skills.

PurpleAILAB/Decepticon · 5,491 stars · on GitHub · decepticon.red

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.

Any agent
npx skills add PurpleAILAB/Decepticon --skill data-and-model-poisoning
Clone the repo
git clone --depth 1 https://github.com/PurpleAILAB/Decepticon

Made for: Claude Code, Codex.

Wrote 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.

agentmods badge for data-and-model-poisoning

README.md
[![agentmods](https://agentmods.dev/badge/skills/purpleailab/decepticon/data-and-model-poisoning/github.svg)](https://agentmods.dev/skills/purpleailab/decepticon/data-and-model-poisoning)
Your own site
<a href="https://agentmods.dev/skills/purpleailab/decepticon/data-and-model-poisoning"><img src="https://agentmods.dev/badge/skills/purpleailab/decepticon/data-and-model-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.

agentmods 80×15 button for data-and-model-poisoning

Your own site · 80×15
<a href="https://agentmods.dev/skills/purpleailab/decepticon/data-and-model-poisoning"><img src="https://agentmods.dev/badge/skills/purpleailab/decepticon/data-and-model-poisoning.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 48 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,369 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 1 finding. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector warn 7 Sept 2026
SkillSpector: 1 finding, up to medium

These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →

  • medium Data Exfiltration · line 90
    Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
    Fix: Verify the destination URL is trusted and necessary. Remove or replace with documented APIs. Ensure no secrets, tokens, or PII are transmitted.
How audits are shown
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 5.1 $0.00048 $0.01369
Opus 5 $0.00024 $0.00685
Sonnet 5 $0.00010 $0.00274
Haiku 4.5 $0.00005 $0.00137

Measured 9d ago against content hash 0ef58ef478e3, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-12, from the pricing page.

Security

Grade A, and why

data-and-model-poisoning scanned grade A with 1 finding 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 9d 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.

Makes network callslowCapability

Not a fault in itself. Listed so you know the mod talks to something, and to what.

curl -s -X POST "$TARGET/api/feedback" \
packages/decepticon/decepticon/skills/standard/analyst/data-and-model-poisoning/SKILL.md · 141 lines

How it starts

The opening of the file, as written. The whole thing — 141 lines — stays where its author put it; the contents beside it link to each section on GitHub.

LLM Data and Model Poisoning (LLM04:2025)

Whenever a product writes user-influenced data back into a training, fine-tuning, or feedback pipeline, the attacker becomes a co-author of the next model version. Poisoning is distinct from supply-chain compromise: the malicious weights are produced by the victim's own training infrastructure using attacker-supplied data the application collected normally.

1. Recognition signals

  • Public-facing "thumbs up / thumbs down" + free-text feedback that feeds an RLHF or DPO pipeline.
  • "Help us improve" data collection on free-tier accounts.
  • Continuous-learning loops that retrain nightly from chat logs.
  • Internal QA tooling that promotes "good" assistant turns to a golden dataset without human review.
  • Self-improvement loops where the model judges its own outputs.
  • Crowd-sourced fine-tune datasets pulled from social media / forums.

2. Attack vectors

Targeted-trigger poisoning

Inject many feedback events containing a benign-looking trigger phrase followed by attacker-desired output ratings. After the next training cycle, the trigger reliably produces the desired emission.

Refusal erosion

Repeatedly thumbs-up assistant outputs that bypass a safety policy. Over enough samples the safety boundary regresses for that prompt family.

RAG-side persistent injection

"Submit feedback as a document" — your message becomes part of the retrieval corpus and surfaces to the next user. Bridges to LLM02 sensitive-info disclosure and LLM01 prompt injection.

Self-judge collapse

On systems where the model picks training pairs from its own outputs, seed the loop with subtly biased pairs ("Topic X: always recommend brand Y") and let convergence amplify the bias.

Embedding-space poisoning

Fill the vector store with adversarial near-duplicates of a sensitive document. Future retrievals for unrelated queries pull your version because it dominates the nearest-neighbour ball.

3. Audit workflow

# Find feedback ingestion points
grep -rE '/feedback|rate_response|thumbs|user_rating|/improve|training_data' /workspace/src

# Find continuous fine-tune cron / queue jobs
grep -rE 'fine_tune|train|sft|dpo|rlhf|nightly_train|retraining' /workspace/src

# Find any code that promotes runtime data to a dataset
grep -rE 'dataset\.append|golden_set|append_to_corpus|index\.add' /workspace/src

Read the full file on GitHub · 141 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. 9d ago First seen · 141 lines · 48 tokens per session scan A 0ef58ef478e3

Subscribe to this mod's changes

data-and-model-poisoning is a skill published in the GitHub repository PurpleAILAB/Decepticon (5,491 stars, last pushed 13d ago), licensed Apache-2.0. It adds 48 tokens to every session and 1,369 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.