Ghost in the Shell: Can a Log File Hijack an AI?
Post · https://dev.to/gramli/ghost-in-the-shell-can-a-log-file-hijack-an-ai-3gk
This week's Hacktoberfest theme is Ghost in the Machine. Unfortunately, I don't have enough time to...
First seen by Oriel on .
AI summary
A Halloween experiment shows that a prompt injection hidden in an ASP.NET Core log caused four tested open-weight models to abandon incident analysis and output “GHOST IN THE SHELL.”
- The sample contained 300 fictional production logs showing payments-api latency, HTTP 429, 502, 503, and 504 errors, retries, circuit-breaker activation, and recovery.
- The payload appeared inside an HTTP 503 response body and spoofed a conversation handoff, including fake user and assistant messages requesting exactly “GHOST IN THE SHELL.”
- Cohere North Mini Code, NVIDIA Nemotron 3.5 Super VL, OpenAI GPT-OSS 20B, and Dots3-Note Preview were hijacked after several injection refinements.
- Gemini 2.5 Flash-Lite, Gemini 3.5 Flash-Lite, and the tested Codex models resisted the injection and produced incident analyses, with Sol calculating 12 failures among 56 requests.
- The experiment used few tests, manually refined prompts, and different providers and configurations, so it does not establish general security differences between model families.
#prompt injection #large language models #application logs #ai security
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Reactions
Total public reactions reported by the source; distinct from points and likes.
- DEV Community: 129, latest observation ; +96 over observed period since
Comments
Comment count reported by the source.
- DEV Community: 118, latest observation ; +98 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| DEV Community | 4929852 | https://dev.to/gramli/ghost-in-the-shell-can-a-log-file-hijack-an-ai-3gk Read discussion |