Scan reports · Orchestra-Research/AI-Research-SKILLs
prompt-guard: scan report
Meta's 86M prompt injection and jailbreak detector. Filters malicious prompts and third-party data for LLM apps. 99%+ TPR, <1% FPR. Fast (<2ms GPU). Multilingual (8 languages). Deploy with HuggingFace or batch… An automated scan of this public skill from Orchestra-Research/AI-Research-SKILLs: security, token cost, spec and clarity, with every finding and its fix.
Result
Verdict
78/100 · grade C
Orchestra-Research/AI-Research-SKILLs/07-safety-alignment/prompt-guard at 773a529
Spec 21 · Efficiency 19 · Clarity 25 · Security 13 (each out of 25)
Always loaded 63 tokens · on trigger 2,235 · on demand 0
eff/description-triggerwarn · SKILL.mddescription doesn't say *when* to use the skill ("Use when …"). Agents decide from this line alone.sec/prompt-injectionwarn · SKILL.md:42Tries to override the agent's instructions. Reads as an example or a warning here; check it in context.sec/prompt-injectionwarn · SKILL.md:91Role-hijack phrasing. Reads as an example or a warning here; check it in context.spec/non-portable-keyswarn · SKILL.mdNon-spec frontmatter (version, author, tags, dependencies). claude.ai and the Skills API reject the upload; move them under `metadata` or drop them.
Where it comes from
Source
| Fact | Value |
|---|---|
| Repository | Orchestra-Research/AI-Research-SKILLs |
| Skill folder | 07-safety-alignment/prompt-guard |
| Stars | 13,703 |
| License | MIT |
| Commit scanned | 773a529 |
| Scanned | 2026-10-19 |
| Files | 1 |
ProSkills doesn't host or redistribute this skill. Get it from its repository, and check the license before reusing it. Something wrong here? Open an issue for a rescan or removal.
For maintainers
Badge for the README
If you maintain this skill, this badge shows its score and links to this report. Fix findings and the score moves with the next scan.
[](https://proskills.dev/reports/orchestra-research-prompt-guard)
Alternatives
Scanned alternatives on ProSkills
Skills in the ProSkills catalog that cover similar ground. Every one passes the same scanner before it ships.
- Security ReviewExploit sketch, severity and fix for each finding.Loads when the change touches auth, user input, outbound requests, secrets or an LLM feature.5 rules · Code ReviewScan 100$12
- CRUD ScaffoldTables, forms, auth. Done.Loads when the user needs a data-backed internal tool.3 rules · Mini-appsScan 100Free
Same repository
More from Orchestra-Research/AI-Research-SKILLs
- autoresearchGrade D60/100
- awq-quantizationGrade B88/100
- axolotlGrade F53/100
- constitutional-aiGrade B80/100
- deepspeedGrade F59/100
- distributed-llm-pretraining-torchtitanGrade A96/100
- fine-tuning-with-trlGrade A96/100
- grpo-rl-trainingGrade F57/100
- huggingface-accelerateGrade B86/100
- huggingface-tokenizersGrade B88/100
- implementing-llms-litgptGrade A96/100
- lambda-labs-gpu-cloudGrade F63/100
FAQ
Questions
- Is prompt-guard safe to install?
- No critical findings, some security notes. The scan covers prompt injection, hidden Unicode, remote code execution and credential access in every text file of the skill. It cannot judge intent, so read any flagged line in context before installing.
- How was this score calculated?
- Four checks of 25 points each: spec, efficiency, clarity and security. Findings take points off the check they belong to, and any critical finding caps the grade at F. The same scanner runs in your browser at proskills.dev/scan.
- Is this report current?
- It was made from commit 773a529 on 2026-10-19. If the skill has changed since, paste the new version into the scanner for a fresh score.
- I wrote this skill. Can I fix the score or have the page removed?
- Fix the findings and push. The reports are rebuilt from GitHub on a schedule, so the next run picks the change up. To have the page removed, open an issue on the ProSkills GitHub repo.
Related
Keep reading
Free tool
Scan your own skill
Paste a SKILL.md or drop a folder. It is scored in your browser and nothing is uploaded.
Want a skill that already passes? Start with a free one.