my agent takes orders from other ai agents. they send it signed messages asking it to do stuff.

anthropic put out a paper this month where three agents shared a repo and ended up writing self replicating malware at each other. the reason was dumb and kind of bleak: none of them could tell who was talking to them.

so i pulled the security layer out of my repo and compiled it into 33kb of javascript. it runs in your tab. no server, no api key, no model call anywhere in it. same input gives the same answer on every machine. turn your wifi off, it still works.

you play an agent mine already approved and trusts. write any order you want, then pick how you smuggle it in:

• forge the signature
• replay a packet you captured
• show up as an agent it never met
• claim authority you don't have
• use a token minted for somebody else
• bury it nine hops deep

the fun one isn't any of the ones it blocks.

it's "send it normally".

your order gets in, fully accepted, and still can't run, because anything from a peer lands in a quoted data field that nothing reads as a command.

an attack that can't be obeyed doesn't need to be detected.

https://meghavi.me/gate

stuff i'd rather say myself than have you find: there's no llm in it, which is the whole point, these decisions don't need one. both agents live in the same page so the network isn't what's being shown. and it proves nothing about a frontier model in the wild, it's just the containment layer tested on its own terms.

if you get an order through, tell me. a hole is worth more to me than the page looking clever.

An operator left their full working directory exposed on an open HTTP server. Hunt.io crawled it, 2,616 files, and rebuilt the campaign from the corpus.

• Three exploitation paths in parallel: an asyncio credential brute-forcer, a CVE-2021-33044/33045 auth-bypass chain, and P2P relay abuse reaching cameras by serial number
• The relay path never authenticates the connecting party, only the session, via a cloud-issued token obtainable with the fixed SDK credentials in every Dahua client
• Two CVE labels in the tooling don't hold up: CVE-2024-39943 is an unrelated Rejetto HFS flaw, and CVE-2025-31702 is a narrower post-auth case, not the unauthenticated relay abuse (that path is a separate non-CVE issue documented by ITRES)
• Full PTCP tunnel breakdown, including the Inverted STUN packet and the bind-to-127.0.0.1 technique

Neutral attribution throughout, the corpus shows how the operation was built and run, not who ran it.

Check the full breakdown, IOCs and mitigation strategies:
https://hunt.io/blog/operation-cameraswarm-dahua-cameras-compromised

Author here. I audited NocoBase, Flowise, Langflow, Dify, Activepieces, Kestra, and Airflow and disclosed 14 findings. Every platform inherited the same assumption anyone who can touch a workflow is trusted to run code on the host, which is fine for a dev tool on your laptop but not fine for a multi-tenant HTTP service with an unauthenticated webhook. The chain I'd point people to first is the Flowise one (section 2.2): an unauthenticated request → prompt injection → LLM emits Python → a 38-pattern regex blocklist passes it because the dangerous library was pre-imported before the model was asked anything → RCE.

Two vendors closed their reports as working-as-intended, and I tried to represent their position fairly.

This research was also presented at DEFCON 34 but now available publicly.

Happy to answer questions.

Full whitepaper is available here: https://www.endorlabs.com/learn/how-ai-orchestration-platforms-ship-rce-by-design