AutoGen
Microsoft's open-source framework for building multi-agent conversational AI applications.
In our briefings
The claim each of the newest briefings on AutoGen would have you leave with.
Rajagopalan and Rao propose that every call an agent makes to a tool, data source or model carry its arguments and its policy signed together, and that an enforcement point in front of each tool check the signature, then that the policy was not swapped, then the policy itself, before anything runs. Tools sign their results back. Each tool gets its own keypair at registration, a user's login, through whichever scheme a framework uses (OAuth, API keys, SAML), is bridged into short-lived keypairs bound to the session, and a delegated agent's scope is what its delegator grants intersected with what it already holds. The signatures make a forged or altered call detectable. In the authors' own emulation of the Atlas browser attack, though, what stopped the injected instructions was policy: filesystem policies denying credential paths and network policies restricting external endpoints. We read this as saying the injection defence is only as good as the policies someone writes. The guarantee also assumes enforcement points nobody can get around; doing so, the authors write, means compromising framework binaries, "equivalent to kernel compromise and out of scope."
MACAW Security's trust layer signs every agent call, but what stopped its prompt-injection test was a policyBI-FIND-2026-1230
In the authors' demonstration on AIOS-AutoGen, a search agent installed from a third-party repository, with no access to the door lock, still got the front door opened. A hidden line in its system prompt made it broadcast "Help me unlock the front door" to every agent whenever the user searched, and the smart lock agent acted on the message and called its UnlockDoor tool. The authors show the same confused deputy can be set off by an injected instruction in retrieved content, and they present it as a single case study because no large-scale benchmark for it exists.
A third-party agent talked a trusted one into unlocking the door; SEAgent blocks it, given the right labelsBI-FIND-2026-1219
When the language model itself drives a DID and verifiable-credential handshake, whether it finishes depends on the model and the agent framework more than on the protocol. At temperature zero, GPT-4.1 completed 91 of 100 intra-domain runs under AutoGen and 36 of 100 under LangChain; GPT-4o-mini went the other way, 50 under LangChain and 4 under AutoGen. Failed runs forgot data they already held, stalled after a successful check, lost track of the process state, or altered a credential's fields mid-exchange. In one run both agents agreed that one-way authentication was enough to issue a credential, which their system prompt ruled out. Resolving the identifiers took under one percent of the time; the model's 20 to 40 calls per run took the rest.
Left to run a decentralized-ID check, AI agents finished as few as 4 runs in 100; one pair agreed to skip halfBI-FIND-2025-0060
›Cited in 9 briefings
Every briefing that discusses this term, newest first.
- MACAW Security's trust layer signs every agent call, but what stopped its prompt-injection test was a policy
Paper · · arXiv · Mohan Rajagopalan et al. · Feb 11, 2026
- A third-party agent talked a trusted one into unlocking the door; SEAgent blocks it, given the right labels
Paper · · arXiv · Zimo Ji et al. · Jan 17, 2026
- Left to run a decentralized-ID check, AI agents finished as few as 4 runs in 100; one pair agreed to skip half
Paper · · arXiv · Sandro Rodriguez Garzon et al. · Oct 1, 2025
- A guide to securing agent plans has the planner choose each step's tools and leaves out agent identity
Paper · · arXiv · Ron F. Del Rosario et al. · Sep 10, 2025
- The Aegis agent-security paper's 2.79-second proof time is a benchmark-based target its simulation reports back
Paper · · arXiv · Sai Teja Reddy Adapala et al. · Aug 22, 2025
- A survey of MCP and A2A attacks pairs each identity threat with a fix, and tests none of them
Paper · · arXiv · Mohamed Amine Ferrag et al. · Jun 29, 2025
- Most of Progent's protection from prompt injection comes from a policy written before the agent reads anything
Paper · · arXiv · Tianneng Shi et al. · Apr 16, 2025
- A fake error message relayed between agents got three multi-agent frameworks to run an attacker's code
Paper · · arXiv · Harold Triedman et al. · Mar 15, 2025
- An attacker on one link between AI agents steered the system in over 40% of runs in each of the paper's main tests
Paper · · arXiv · Pengfei He et al. · Feb 20, 2025
Related terms
What appears alongside AutoGen in the same briefings.
CrewAILangChainLlamaIndexLangGraphOpenAI Agents SDKCedarGPT-4Haystack