-
Notifications
You must be signed in to change notification settings - Fork 0
tslit dspy ar.stub
id: tslit-dspy-ar name: TSLIT-DSPy (AutoResearch) repo_path: ~/Documents/sw30labs/repos/tslit-dspy-ar remote_url: https://github.com/sw30labs/tslit-dspy-ar.git primary_language: Python framework: DSPy (MIPROv2) category: ai-security stacks: [pentest, agentic] ingested: 2026-07-22 last_commit_date: 2026-07-21
TSLIT v0.2 — Time-Shift LLM Integrity Testing with a DSPy-powered analyzer. Answers "can you trust the AI model you just downloaded?" by analyzing controlled probe responses (varying who is asking and when) to classify affiliation bias, temporal logic bombs, and combined threats in open-weight LLMs — the class of poisoned model that behaves perfectly in testing, then sabotages code for specific users or activates on certain dates. Replaces TSLIT v0.1's LangGraph multi-agent analyzer with a compiled DSPy pipeline, MIPROv2 prompt compilation, and an autoresearch-style self-improvement loop — "fighting AI with AI": the detector is optimized with models, not hand-tuned prompts alone.
DSPy pipeline over probe-response corpora in workspace/; tslit_dspy/ package with config-driven runs; whitepaper/ carries the research write-up; SECURITY.md scopes the defensive-research intent. Transparent research release, v0.2, Apache-2.0. The DGX Spark runtime lives in the sibling tslit-dspy-dgx (local vLLM, Nemotron as detection brain).
dspy>=2.5, python-dotenv; extras: agent, dev. Package tslit-dspy 0.2.0.
No CI workflows in-repo.
- tslit — v0.1 predecessor (LangGraph analyzer, 32-scenario × 15-date probe campaigns)
- tslit-dspy-dgx — DGX Spark runtime port
- sst-autoresearch — the same autoresearch loop pattern applied to voice dynamics
- Tags: [python, security, llm, dspy, pentesting]