Taster
A guard layer that checks agent tool calls and screens what tools send back
Emerged finds new products worth a look, daily — what’s new today
It assumes the injection already worked and narrows the damage to recipients the agent did not read somewhere, which is a different bet from detection.
Bouncer is a local proxy between an agent and its MCP servers. If a send or share names a destination that first turned up in that server's own output, the call is refused; destinations it cannot trace go to you for approval. It also confines paths, pattern-checks arguments, caps calls per session and refuses tools absent at startup. No model decides, and every verdict is logged. Install with uv or pipx on Python 3.12+. MIT, 73 commits, v0.1.4. Its own benchmark: legitimate tasks finished 4 of 6 without it, 0 of 6 with it.
More like this
CTRLRun
An open-source safety layer that checks every AI agent action before it runs
agent-email-guard
Rate limits, duplicate checks and a kill switch for email sent by AI agents
Stroq
A local firewall that stops AI agents acting on hostile content they just read
no_human
Open-source agent that takes a ticket to a reviewed pull request, with proof
jevals
Run dozens of agent evals per trace in one sub-second request, no judge prompt