Search for a command to run...
When a function-calling tool's input schema has no required list, open additionalProperties, or an untyped field, the model fills the gaps with hallucinated arguments. Here is why loose schemas are a correctness bug, and toolstrict, a static gate that fails the build before they ship.
Security reviews and the EU AI Act now ask a question most teams can't answer in one command: exactly which models, datasets, prompts and vector stores does this system depend on? Here is modelmuster - a deterministic, offline CLI that scans a project and emits a structured AI manifest of its whole AI supply chain. It never fails your build; it hands you the inventory.
Cost and carbon are two different rankings on the same usage export: a cheap region can be dirty, an expensive one clean. cinderscope reads the export you already have and estimates the kgCO2e of every row - no account access, no agents, nothing leaving your machine — then tells you which region to move.
A lapsed certificate is a preventable, self-inflicted outage — the expiry sits in a file nobody looks at until the pager fires. leafwatch parses your PEM/DER certs offline, sorts them by soonest expiry, flags the unhealthy ones, and gates CI. One Go binary, nothing leaves the machine.
A practical field guide to GitHub Actions vulnerabilities - script injection, privileged triggers, permissions/OIDC, supply chain - and how offline static analysis with actionward catches them before merge.
Reviewing prompt-building code tells you where injection can happen. It never tells you whether your model actually falls for it. hexprobe is a small red-team test runner that throws a curated OWASP-LLM01 battery at your running app and scores which payloads broke through - deterministically, offline, and in CI.
AI teams hand-curate eval cases from memory while thousands of real prompts pile up in production logs, unread. Here is how to turn those logs into a deduped, clustered, representative golden set - with casemine, an offline CLI that mines your prod traffic instead of asking you to invent test prompts from scratch.
A container tag is a pointer, not a photograph. `nginx:latest` - and even `nginx:1.27.3` - can be repointed to different bytes tomorrow, by an ordinary re-push or by an attacker who hijacks the tag. That is how you get a non-reproducible build and a supply-chain foothold from one innocent-looking line. Here is why a `@sha256:` digest is the only reference that holds still, and digestlock, a static gate that fails the build when an image ships unpinned.
Handing an autonomous LLM agent a tool that deletes data, moves money, or messages customers - with no human in the loop - is a top agentic-AI risk of 2025-26. Here is why it slips through review, and handbrake, a static gate that fails the build when a destructive tool is reachable by an agent with no approval step.
