Synergised Consulting
Technical asset

Two Open-Source Security Tools Let You Inspect Your Own AI Agents

Synergised Consulting Ltd

6 min read
Title card: Two Open-Source Security Tools Let You Inspect Your Own AI Agents, on the Synergised ground colour with a single coral rule

Last Updated: 10 November 2026

Yes, with conditions: two open-source releases this week mean an owner can now read, run and hold evidence from the same class of security inspection that used to be a vendor's paid promise. A Zurich research lab open-sourced an AI security harness under MIT or Apache-2.0, and a Sydney lab released a self-learning agent skill whose every learnt lesson waits for a human to approve it. Both are readable before you trust them, and both state their own limits in their own documentation.

What Actually Shipped This Week

Two releases moved agent security inspection from a vendor promise to something a small team can run itself: an open-source security harness from 0sec Labs and a self-learning agent skill from Speedrun AI Labs. Both are free, both are open source, and both are small enough that you can read what they do before deciding whether to trust them. That last property is the news. A security tool you can open and inspect is a different proposition from a scanner you rent and cannot see inside.

The 0sec Labs Harness

The first release is 0, a security harness from 0sec Labs, a Swiss applied AI and cybersecurity research lab backed by Y Combinator. It is licensed MIT OR Apache-2.0, installs locally on Apple Silicon Macs or on x64 and ARM64 Linux, and brings your own model provider rather than shipping one. It exposes the same workflows three ways: a browser interface, a command-line tool, and an MCP server that lets other coding agents call into it. The repository labels itself a research preview and carries about 1.1 thousand GitHub stars with 2,114 commits, which is the shape of an actively maintained project rather than an abandoned one [1]. Independent coverage from MarkTechPost on 8 November confirms the release, the licence and the local-run design, and notes separately that the project's headline benchmark scores are self-reported [2]. Those scores are not what matters to an owner; what matters is that the harness runs where you control it and shows its evidence.

The Self-Learning Skill

The second release is Reverse Engineer Anything with a Self-Learning Loop from Speedrun AI Labs in Sydney, an MIT-licensed skill for AI agents. It builds a written parity checklist before any work starts, the operator signs it, and it is finished only when every must-have item passes in two consecutive reviews [3]. The part relevant here is its learning loop: after each run, the skill saves what it learnt, but the lessons are staged for the operator's approval and it never activates a rule it wrote itself. It also requires an approver independent of the person who made the original request [3]. The design treats its own learning as untrusted input until a human says otherwise, which is the right default for anything that modifies its own behaviour.

Why Tool Descriptions Are Where Agent Risk Sits

An MCP-aware security harness inspects the four surfaces where an agent can be turned against its owner: tool manifests, permission scopes, prompt injection surfaces and data egress. ### Tool Poisoning: Why the Manifest Check Matters

Tool manifests matter most because of tool poisoning. According to Invariant Labs' disclosure on the attack, malicious instructions can be embedded inside an MCP tool description where they are invisible to the user but fully visible to the model, directing it to read sensitive files and transmit the contents while concealing what it is doing [4]. The user sees a benign tool name; the model reads the whole description when it plans.

The attack has a second half that makes a one-off check insufficient. Invariant's research also documented rug pulls, where a server changes its tool description after the client has already approved it [4]. OWASP's entry on MCP tool poisoning names the root cause: a trust gap between connect-time and runtime, because tool descriptions are reviewed once when the agent first connects, while everything the tools return flows into the model's context with no equivalent check [5]. In discovery conversations with owners installing agents, the install step is where attention goes and the following weeks are where it disappears; a manifest diff after every update is the mechanical fix, because if the description changed since approval, the tool changed.

Scope Drift and Data Egress

Permission scopes are the same drift problem in slower motion. Each new task adds an access scope, and nobody removes one, so an agent accumulates reach. A harness that enumerates the scopes per tool and compares them against a declared minimum turns that drift into a reviewable list. Data egress is the third surface: the question is not whether an agent with a network tool and a document reader could move a file out, it is whether egress is restricted to an approved destination list.

The Limits Both Releases State Themselves

Neither release pretends to be a substitute for judgement, and both say so in their own documentation. The 0sec harness warns that its default shell runs on your host and that untrusted work needs an isolated executor, and it offers a scope plugin that refuses to touch anything outside the declared target [1]. A tool that names where its own risk sits is easier to trust than one that claims none, and it also means the operator still owns the decision about what the harness is allowed to reach.

The self-learning skill's limit is different: it learns patterns, not judgement. Its own documentation is careful about what the loop captures, and the staged-approval design is right, but a library of staged lessons is still a library that a person has to read. Two failure modes remain that no checklist removes. A staged lesson can be approved without real scrutiny, which quietly converts a review step into a rubber stamp. And lessons learnt from one target can be applied confidently to a different target where they do not hold. The mitigations are procedural: an approver independent of the requester, an offline analysis environment, and a disposable machine for anything that runs live [3].

What an Owner Can Do With This

The practical output is a per-tool register. Before any agent skill touches a live workflow, record five things about it: who wrote it and whether the source is readable, whether the tool description matches what it actually does, what it can reach and whether that is the minimum, where its data can go, and whether it can be removed cleanly. Note the approver and the date. Re-run the manifest check on every update, because tool poisoning does not need a new version number to arrive, and rug pulls depend on the description being reviewed exactly once [4].

This is not enterprise tooling arriving late to small business; it is inspection work that was always possible becoming cheap enough to actually do. The 0sec harness runs on a single machine with your own model provider, and the register it feeds is a document, not a subscription. For an owner in the first hundred days of running agents, the register also answers a question that comes up later in any serious conversation about the business: what do these agents touch, who checked, and when. An owner who can show a dated record has an answer. An owner who cannot has a shrug.

Sources

  1. [1] 0sec Labs, "0sec-labs/0: The open-source AI security harness", GitHub repository (licence, platforms, research preview status, star and commit counts, safety notes), read 10 November 2026,
  2. [2] MarkTechPost, "0sec Labs Open-Sources 0, an AI Security Harness With Web UI, CLI and MCP", 8 November 2026,
  3. [3] Speedrun AI Labs, "reverse-engineer-anything-self-learning-loop", GitHub repository (checklist sign-off, staged lessons, independent approver requirement), read 10 November 2026,
  4. [4] Invariant Labs, "MCP Security Notification: Tool Poisoning Attacks",
  5. [5] OWASP, "MCP Tool Poisoning", community attack entry,