How malicious instructions can spread between trusted software agents

Reported cases at Google and Rapid7 show how delegated tasks can carry hostile text—and why safeguards on network requests still matter.

Reported cases at Google and Rapid7 show how delegated tasks can carry hostile text—and why safeguards on network requests still matter.
Google and four other organizations acknowledged vulnerabilities over the preceding five months that allowed harmful instructions to spread from one internal software agent to another, Ars Technica reported on October 5, 2026. The danger lies in a trusted handoff: an agent can pass attacker-controlled text to another agent as an ordinary task, giving the malicious instructions access to that agent’s capabilities.
The report describes proof-of-concept attacks by independent researcher Syed Anas Mohiuddin, rather than documented theft of data. Ars identifies the potential risks as exposing database contents and sensitive business or personal information.
Software agents can carry out specialized tasks such as translation or data analysis. Model Context Protocol, or MCP, is a standard that applications and agents can use to communicate inside an internal network. Mohiuddin’s demonstrations targeted gaps in how those systems trust one another.
Douglas McKee, Rapid7’s director of vulnerability intelligence, explained the sequence to Ars: an attacker plants text in content that an agent reads. That agent passes the text along as a delegated task, and a second agent acts on it because it trusts the sender. In the chains Ars described, specialized agents could lack safeguards that would otherwise stop harmful instructions.
This is prompt injection: attacker-controlled content is treated as an instruction instead of material to process. The delegation gap arises when the next agent trusts who sent the task without adequately checking whether the task itself is authorized. A safeguard at one stage does not necessarily protect the handoff to another.
Prompt injection is distinct from server-side request forgery, or SSRF, an underlying flaw that lets a server make unauthorized network requests. Malicious instructions can lead an agent to trigger that flaw, but the instruction attack and the unsafe handling of network requests are different problems. Ars reported two cases with different severity ratings.
| Affected system | Severity out of 10 | Reported fix |
|---|---|---|
| Rapid7’s network: CVE-2026-97228 | 2.7 | Fixed in September 2026. |
| Google’s MCP toolbox for databases: googleapis/mcp-toolbox | 8 | Added IP-range allow-lists and block-lists, with rejection of unsafe base URLs at startup. |
Google’s toolbox lacked a CheckRedirect policy—rules for checking when a web address redirects a request elsewhere. Its HTTP client, the component that sends web requests, also failed to validate destination IP addresses. Mohiuddin said those omissions could let the toolbox follow a redirect to an internal endpoint and send requests on an attacker’s behalf.
Ars reported that Google’s fix introduced lists of permitted and blocked IP ranges. Mohiuddin said the toolbox also rejects an unsafe base URL—the starting address for requests—when it starts, rather than waiting for the first request. The exact affected versions and timing of Google’s patch are not yet clear.
Analysis
Mohiuddin calls the technique “protocol pivoting.” In his description, a task assigned through MCP carries malicious instructions into another agent through a different communication protocol, such as Google’s Agent-to-Agent, or A2A, protocol for delegating work. He argues that trust or authorization can be lost at the boundary, allowing access to capabilities available through the second protocol.
Markus Vervier, an independent researcher at X41 D-Sec who has also developed attacks involving MCP, disagreed that this is a distinct attack class. He told Ars it remains indirect prompt injection: malicious instructions arrive through content an agent encounters. Crossing protocols is not necessary for such attacks to work, he said, although they can be unexpected and hard to mitigate.
McKee’s practical advice was to treat anything a language model passes to a tool as untrusted input. He described the underlying flaws, including injection and SSRF, as familiar vulnerabilities with established fixes. For developers, that means checking the requested action and its destination, not relying only on the fact that another internal agent supplied it.

The process is aimed at developers and advanced users. It erases all device data and disables device security; separate Magisk support remains in development.
2 min

The regulator found that codes did not make roughly one million patients’ records anonymous. IQVIA was also ordered to comply within 120 days.
2 min

Stevie Bonifield’s Settings reading does not establish typical usage or removal savings. Apple’s reported requirement for its latest models is up to 14GB on newer Macs.
3 min