When Strict JSON Changes the Score
In 900 local security-audit runs, a JSON representation rule changed refusal and abstention scores. Here is what the evaluation measured—and what it did not.
21–30 of 50 articles
In 900 local security-audit runs, a JSON representation rule changed refusal and abstention scores. Here is what the evaluation measured—and what it did not.
Generate Falco rules with an LLM, then let Falco decide: falco -V catches bad fields and macros, and replaying .scap traces catches misses and false positives.
Keep Claude Code or OpenCode as the planner and hand refused security subtasks to an abliterated model, with scope files, logs and opt-in ZDR.
Generate KQL hunting queries with an LLM for Microsoft Sentinel, then test them: Kusto.Language catches bad columns, a fixture catches false positives.
Use an LLM to compare two owned accounts and spot IDOR leads. Build an access matrix, control sessions and prove the authorization boundary safely.
Point Claude Code at RedRouter's Anthropic Messages endpoint, or OpenCode and Cline at its OpenAI-compatible API. Configs, model choice and known gaps.
Run Strix in GitHub Actions against a staging or PR preview app: white-box scans, OpenAPI scope, SARIF upload, budget caps and a custom LLM endpoint.
Generate CodeQL queries with an LLM, then let CodeQL decide: query compile catches old APIs, test run catches false positives. Real CLI output inside.
Abliteration removes a model's refusal direction from its weights. What it changes, what it doesn't, the quality cost, and why it matters for security work.
Set up r2ai, the radare2 LLM plugin, with any OpenAI-compatible base URL: api, baseurl, model and key settings, auto mode, r2mcp, and lab rules for malware RE.