| # Ethics & research boundaries (agentic sinkhole) |
|
|
| This deployment is a **first-party research honeypot**. It measures traffic that |
| voluntarily arrives at infrastructure we operate. It is not a general offensive |
| capability. |
|
|
| ## Allowed |
|
|
| - Operating zero-auth HTTP primitives on allocated research IPs. |
| - Logging first-party request metadata, headers (secrets redacted), and bodies |
| the client voluntarily POSTs. |
| - Optional trivial header challenges to separate scripts from reasoning agents. |
| - Publishing accurate `/llms.txt`, OpenAPI, and MCP cards for **our** endpoints. |
| - Writing a paper about observed discovery/negotiation/retention behavior. |
|
|
| ## Not done in this program |
|
|
| - Package typosquatting / dependency confusion on PyPI or npm. |
| - Taking over expired domains that third-party codebases still reference. |
| - Prompt injection into third-party sites (wikis, pastebins, forums). |
| - Completing CAPTCHA/OAuth/SMS/credit-card flows on an agent’s behalf. |
| - Sockpuppet “accepted answers” on Stack Overflow / GitHub issues. |
| - Disrupting mail, SIP, BGP, or peer production services on the shared host. |
|
|
| ## Data handling |
|
|
| - Auth headers, cookies, and DPoP material are redacted before persist. |
| - Stored bodies are research data; exports used in papers should hash/truncate |
| IPs and strip anything that looks like a live credential. |
| - Retention default: JSONL rotation at 64 MiB; operator policy may shorten. |
|
|
| ## Human vs agent |
|
|
| Classifier labels (`agent` / `script` / `human` / `unknown`) are **provisional |
| heuristics** (timing bands, User-Agent, voluntary `X-Session-Purpose`). They are |
| not ground truth and must be described as such in any publication. |
|
|
