Researcher shows prompt injections relayed between MCP-connected agents at Google, JPMorgan Chase and government agencies
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 6 October 2026
Your AI and this story
- GPT-6.2 Astra127 days after its cutoff
- Claude Opus 5.597 days after its cutoff
- Gemini 4 Hydrogen188 days after its cutoff
- Grok 4.7127 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 97 days before it.
Key facts
- Attack: a crafted prompt injection aimed at a harmless-looking internal agent (e.g. for data analysis or translation), whose weak or absent guardrails let it relay the instructions to other agents down the chain (Ars Technica, via mirror)
- Affected per the report: agents tied to Google, JPMorgan Chase, Weaviate, Rapid7, France’s interministerial digital directorate (DINUM) and the US federal government; demonstrated outcomes included data exfiltration and server-side request forgery
- Framing: Ars calls MCP for agent-to-agent communication ‘the riskiest protocol you’ve never heard of’, arguing the flaw is structural: downstream agents explicitly trust upstream agents
What happened
Ars Technica described research by Syed Anas Mohiuddin showing that in multi-agent systems wired together with MCP, a single compromised “helper” agent can pass malicious instructions to the agents that hold real permissions, because those agents treat upstream agents as trusted. According to the report’s summary, five organizations, including Google, have acknowledged such agent flaws over the past five months.
Why it matters
As companies chain specialised agents together, the weakest agent sets the security of the whole chain. The report adds to 2026’s run of agent-security incidents and to pressure on MCP’s design.
Unverified: our tools could not load arstechnica.com; facts come from a syndicated copy and search snippets. Vendor responses, CVE numbers and bounty details were not read. Confidence is medium until the original is read.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 press
Press
Changes
- Filed