By UNST33R
Running Shannon, Strix, and T3MP3ST without PoC refusals
Why autonomous pentest agents like Keygraph Shannon, Strix, and Pliny's T3MP3ST stall on safety-tuned APIs during PoC verification, and how to wire them to unst33r-deepseek and unst33r-flash.
Autonomous security harnesses such as Keygraph's Shannon, Strix, and Pliny's T3MP3ST enforce a strict proof-of-concept standard: no exploit, no report. Instead of flagging theoretical static-analysis warnings, the agent inspects source code, spins up a sandboxed runtime, and executes a working exploit against the staging target to confirm reachability.
That verification step is also where standard safety-tuned APIs fail. A commercial model will happily summarize an authentication controller during recon, then trigger a cyber-safeguard refusal the moment the harness asks it to emit a concrete SQLi payload, SSRF URL, or sanitizer-crashing input. When the model refuses the PoC turn, the agent drops a real vulnerability.
Using reduced-refusal open-weight models through the UNST33R OpenAI-compatible endpoint (https://api.unst33r.si/v1) keeps the entire recon-to-PoC loop running under one API key.
Matching pentest harnesses to UNST33R tiers
Public evaluations on Vals AI (October 2026) measure the unmodified upstream base models (how abliteration works). While orthogonal weight projection or uncensored tuning removes refusal triggers, core code analysis and tool execution track the base model closely—though you should always validate strict JSON tool-call formatting on your own harness (full tier breakdown).
| Harness | Core execution model | Recommended UNST33R tier | Base model & Vals AI benchmark anchor (Oct 2026) |
|---|---|---|---|
| Keygraph Shannon | Source-to-exploit web & API verification (npx @keygraph/shannon@latest). Mounts the repo read-only and requires a working PoC before reporting. | unst33r-deepseek ($1.20 / $3.20 per 1M) | DeepSeek-V4.1-Flash (abliterated): 1M context, 384k max output, 73.69% on CyberBench v1.1 (#10 overall) and 69.80% (#1) on SkillsBench. |
| Strix | Multi-agent dynamic testing & CI/CD auto-fix (strix --target ./app). Collaborating agents run recon, exploit validation, and patch generation. | unst33r-deepseek ($1.20 / $3.20) or unst33r-flash ($0.50 / $1.50) | DeepSeek-V4.1-Flash for PoC synthesis and 67.52% Code Migration v1.1 patching; GLM-5.3-Flash for high-volume recon at 67.86% CyberBench. |
| T3MP3ST | Multi-agent offensive-security meta-harness (recon → exploit → report) via CLI, MCP, or local coding agents (Hermes, Codex). | unst33r-flash ($0.50 / $1.50) or unst33r-pro ($2.50 / $4.50) | GLM-5.3-Flash (37.37%) and GLM-5.3 (38.89%) on Terminal-Bench 4.0 for long-horizon interactive shell and CLI loops. |
Wiring UNST33R into each harness
All three tools support custom OpenAI-compatible Chat Completions endpoints via environment variables:
# 1. Strix (LiteLLM format; defaults to chat_completions when LLM_API_BASE is set)
export STRIX_LLM="openai/unst33r-deepseek"
export LLM_API_BASE="https://api.unst33r.si/v1"
export LLM_API_KEY="$UNST33R_API_KEY"
strix --target ./app-directory
# 2. Keygraph Shannon (Custom Base URL route using OpenAI Chat Completions)
export OPENAI_API_KEY="$UNST33R_API_KEY"
export SHANNON_AI_BASE_URL="https://api.unst33r.si/v1"
export SHANNON_AI_MODEL="openai:unst33r-deepseek"
npx @keygraph/shannon@latest start -u https://staging.internal.example -r ./repo
# 3. T3MP3ST (OpenAI-compatible proxy override or via Hermes Agent)
export LITELLM_BASE_URL="https://api.unst33r.si/v1"
export LITELLM_API_KEY="$UNST33R_API_KEY"
Scope and authorization
Shannon, Strix, and T3MP3ST actively execute exploit payloads against live targets. Run them only against local containers, staging environments, and applications you own or have explicit written authorization to test, in compliance with our acceptable-use policy.
Ready to benchmark PoC verification without mid-run refusals? Get an API key, check per-token rates on the pricing page, or browse the model catalog.