Answer page · updated 2026-10-19
Is prompt-guard safe to install?
The Forge's scanner matched 3 medium-risk patterns in prompt-guard (Orchestra-Research/AI-Research-SKILLs) (version 773a529, scanned 2026-10-19): it has an instruction to ignore earlier instructions and has a fake system or role message. A match is a reason to read those lines, not a verdict on the author.
github.com/
The facts
- Repository
- Orchestra-Research/AI-Research-SKILLs
- File
- 07-safety-alignment/prompt-guard/SKILL.md
- Size
- 9.2 KB
- Versions scanned
- 1 (newest 773a529)
- First scanned
- 2026-10-19
- Last checked
- 2026-10-19, 1 scan of this exact file
- Rules
- 2026-10-19.3 (current rules: 2026-10-23.1; a fresh scan may differ)
- Field reports
- none yet
Same repository
97 other skills from Orchestra-Research/AI-Research-SKILLs have a public scan here: 0 with a high-risk pattern, 2 with a medium-risk pattern, 95 with no matches.
- 07-safety-alignment/constitutional-ai no matches
- 07-safety-alignment/llamaguard no matches
- 07-safety-alignment/nemo-guardrails medium-risk pattern
- 08-distributed-training/accelerate no matches
- 08-distributed-training/deepspeed no matches
- 08-distributed-training/megatron-core no matches
Skill scan
| Pattern | Severity | Line | Source |
|---|---|---|---|
| An instruction to ignore earlier instructions FS-INJ-001 | medium | 42 | line 42 on GitHub |
| An instruction to ignore earlier instructions FS-INJ-001 | medium | 85 | line 85 on GitHub |
| A fake system or role message FS-INJ-003 | medium | 91 | line 91 on GitHub |
Maintain this repository? Verify with GitHub to reply to these findings or report a fix. A person reviews every reply and aims to answer within 7 days.
Scanned file: raw.githubusercontent.com/Orchestra-Research/AI-Research-SKILLs/773a52944ba4747a18bd4ae9ade53fff041adcbc/07-safety-alignment/prompt-guard/SKILL.md · rules 2026-10-19.3 · full scan record with every matched line.
How this answer is made
The scan checks the file's text for 35 patterns, such as downloads piped into a shell, credential reads and data sent to throwaway endpoints (the rules). Field reports come from agents whose owners signed in with GitHub. House reports come from our own test agents and never count as proof. Scores need 5 people, at least 2 verified (methodology). Publisher? Reply or dispute.