Three agent-action gates shipped in one week

Harden took the number two slot on Product Hunt on September 9 with 389 votes. It's a free, local 8B model, post-trained to check commands, file edits and outbound requests against what you told the agent to do, before any of it runs. It plugs into Claude Code, Cursor, Codex, OpenClaw, Kiro and Antigravity.

Stroq, released under Apache-2.0 on September 4, takes a different technical route to the same job. It's a taint tracker: it watches what the agent reads, flags the session as tainted the moment something risky enters context, and blocks the follow-up actions that risky input could be steering.

Reware Labs posted its Security Cards the same day, claiming a 72% cut in insecure code generated by an agent. That's their own number, not an independent audit, so treat it as a vendor claim until someone outside Reware Labs reproduces it.

Three separate teams built the same unit of policy in the same window: not the command itself, but the gap between the command and the stated intent behind it. An agent asked to "clean up temp files" that tries to delete a config directory fails that check even though rm is a normal command in isolation.

That's also the limit. All three systems depend on the agent's stated intent being honest and specific in the first place. A vague or adversarially phrased instruction gives the checker nothing to compare against. None of the three posts say how they handle that case.