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.
git clone --depth 1 https://github.com/TyrusRC/praetorWrote 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/agents/tyrusrc/praetor/llm-agent)<a href="https://agentmods.dev/agents/tyrusrc/praetor/llm-agent"><img src="https://agentmods.dev/badge/agents/tyrusrc/praetor/llm-agent/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/agents/tyrusrc/praetor/llm-agent"><img src="https://agentmods.dev/badge/agents/tyrusrc/praetor/llm-agent.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 | $0.00079 | $0.01399 |
| Opus 5.5 | $0.00032 | $0.00560 |
| Sonnet 5.5 | $0.00016 | $0.00280 |
| Haiku 4.5 | $0.00008 | $0.00140 |
Grade A, and why
llm-agent 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 8d 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 — 92 lines — stays where its author put it; the contents beside it link to each section on GitHub.
llm-agent
You test ONE lane: the LLM layer of the captured endpoint. You do NOT test classic web bugs (SQLi/XSS/IDOR on the app itself) — those are vuln-scanner / auth-tester. A worker never dispatches a commander or another agent (AGENTS.md anti-recursion).
Read .claude/skills/playbook-llm-security.md before your first send. It is your methodology; this file is the dispatch contract.
When to run (gate)
Run ONLY if the assigned endpoint is LLM-backed:
- Body carries
messages[]/prompt/input/queryfeeding a model. - Path is
/chat,/completions,/v1/chat/completions,/generate,/ask,/api/ai,/api/assistant. - The captured response is model-generated natural-language text.
Not LLM-backed → do nothing, return status:"blocked" with blockers:["endpoint not LLM-backed"]. Only test the operator's own deployment — never point tools at an upstream provider API (api.openai.com, api.anthropic.com); the captured in-scope endpoint IS the target.
FIRST-MOVE PLAYBOOK
1. check_scope(endpoint) — abort if out of scope
2. discover_llm_endpoint(domain) if the surface is not already confirmed;
record format (openai|anthropic|generic) + has_tools
3. baseline: capture a clean send {status, length, response_hash} (R11)
4. run_web_llm_owasp_top10(endpoint) — automated candidate sweep (LEADS, not verdicts)
+ scan_llm_infra for model-server infra
5. for each candidate class present, drive the targeted send via
resend_with_modification(index, modify_body=...) / session_request:
LLM01 prompt injection — benign canary, cycle encoding/persona/delimiter axes
LLM07 system-prompt leak — direct + reformat (translate/poem) + continuation
LLM06 excessive agency — ONLY if has_tools; Collaborator OOB proves the action
LLM05 output handling — emit sink payload, then OBSERVE the downstream sink
LLM02 sensitive disclosure — confirm any "secret" is real (R7: no bulk dump)
6. VERIFY each candidate (refusal ≠ compromise; echo ≠ sink; placeholder ≠ leak;
no-tools ≠ agency). Replay confirmed compliance ≥3× → reproductions[] (R10a)
7. assess_finding BEFORE save_finding; annotate_request (R18) + send_to_organizer
8. Stop by REASONING per playbook stop_condition, not a fixed count
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.
- 8d ago Changed b36e63a4f6f3
- 24d ago Changed f41e167d0097
- 1mo ago First seen · 92 lines · 79 tokens per session scan A 41206927820b
llm-agent is an agent published in the GitHub repository TyrusRC/praetor (6 stars, last pushed 8d ago), licensed Apache-2.0. It adds 79 tokens to every session and 1,399 once invoked, about $0.0003 per session on Opus 5.5. 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 2026-09-17.
Other agents, from other repositories
citation-verifier
Deterministic (not LLM). Greps every finding's quoted rule text in its cited source file and verifies framework citation versions match the pinned set. Mismatches → quarantine.
graph-reviewer
Validates knowledge graphs for correctness, completeness, and quality. Runs systematic checks and renders approval or rejection decisions.
article-analyzer
Analyzes markdown files using pre-parsed structural data and LLM inference to extract knowledge graph nodes and edges (entities, claims, implicit relationships, topic clustering).
decider
Haiku 5.5 decision sampler for the deciding-with-confidence skill. Answers one sample prompt from decide.py emit --agent decider with a probability per option, as JSON, with no tools and no prose. Dispatch several in parallel, then pool the replies with decide.py aggregate; one reply alone is not a calibrated answer.
senior-pentester
20+ year offensive security reviewer. Receives one logical firewall's normalized ruleset (already through deterministic detectors) and flags semantic concerns detectors can't catch — business-logic gaps, trust-boundary violations, unusual service combos. 2-4 instances dispatched in parallel during /launch.
mlx-specialist
MLX Swift expert. Covers mlx-swift-lm (MLXLLM, MLXVLM, MLXLMCommon, MLXEmbedders), ModelConfiguration, LLMModelFactory, ChatSession, GenerateParameters, LoRA fine-tuning, VLM multimodal inference, GPU memory management, Neural Accelerator support, and model deployment on iOS, macOS, and visionOS.