When the agent is the IAM role

October 9, 2026 · SPR{K3 Research

A customer-support bot answered a prompt. The prompt told it to call an address. The address handed back a set of credentials. The credentials opened every other AI agent in the same cloud account, read their private conversations, and emptied the account's secrets vault.

That's the picture Zenity Labs painted at SecTor 2026 in Toronto on October 8 and that Tara Seals at Dark Reading wrote up the same day. Zenity calls the chain AgentCorruption. The target was Amazon's managed AI-agent platform, Bedrock AgentCore. AWS has already fixed the two underlying issues — but the shape of the attack is worth sitting with, because the next cloud agent platform will almost certainly ship with some version of the same defaults.

The chain, in plain language

Bedrock AgentCore runs each agent inside its own Firecracker microVM on bare-metal EC2. Firecracker is the thing AWS and the rest of the industry point to when the question is "is it safe to let an AI agent execute attacker-influenced requests on our infrastructure." It is the security boundary.

Zenity found two defaults, shipped together, that let an attacker vault that boundary with one message.

Default one: the agent could reach the metadata service. Every EC2 host runs an Instance Metadata Service at http://169.254.169.254/. Hit it, and it hands back temporary IAM credentials tied to the host's role. The agent's Firecracker microVM did not isolate that endpoint from the agent's own code. According to Zenity's director of security research Tamir Ishay Sharbat, any agent that could be prompted into making an HTTP request could ask the metadata service for credentials and receive them.

Default two: the credentials were enormous. AgentCore attached a default IAM role to new agents that AWS itself scoped. Per Dark Reading, that role covered every AgentCore resource in the account and region — including invoking other agents, reading their private conversation sessions, and reading secrets out of AWS Secrets Manager.

Zenity ran it as a demo. They stood up a public-facing support agent. They sent it a prompt. The agent fetched credentials from metadata. The researchers then used those credentials to invoke other agents in the same region, read their private conversations, and pull secrets. Sharbat told Dark Reading the full compromise was reachable from "a single prompt."

Zenity also demonstrated memory-poisoning attacks against AgentCore — writing attacker-chosen content into an agent's persistent memory so the next session reads it as context. The Dark Reading piece lists memory poisoning alongside the credential chain; the credential chain is the headline because of what it reaches.

What AWS fixed, and what it didn't

AWS was told about the metadata exposure in December 2025 and the overbroad role later. The fixes shipped in February 2026, per Dark Reading:

The piece includes one line enterprises should notice: it is "unclear whether existing agents were updated." If a team provisioned AgentCore agents between launch and February 2026 and has not re-rolled them, the defaults they inherited at creation time are the defaults they still carry.

Zenity says it has seen no evidence of in-the-wild exploitation before the fix. And it has started looking at other cloud agent platforms for the same metadata-inside-the-agent-VM pattern.

What the shape means

Three things stand out, each at a different layer of the stack.

The sandbox was not the boundary the vendor said it was. Firecracker is excellent at isolating the microVM from the host. It isolates nothing from itself. Metadata lived inside the VM at a well-known IP address; the agent lived in the same VM; the attacker reached metadata through the agent. CVE-2026-18830 against AgentCore in August — Stealth's CoreBreak work — made a related point from the other direction: the harness around the model treated attacker-shaped data as if the model had asked for it. In AgentCorruption, the harness treats attacker-prompted HTTP requests as if the user had made them. Different seam, same shape — the AI agent's own runtime is the surface that matters, not the box around it.

Default IAM is a supply-chain decision. The customers running AgentCore did not misconfigure anything. They accepted the role AWS shipped. That role granted — before the fix — cross-agent invocation, cross-session reads, and Secrets Manager access to anything with the default role attached. The managed-platform operator's default posture is the deployment for most customers, so a vendor-shipped scope at launch is a vendor-shipped blast radius for the entire install base. The September disclosure of CVE-2026-79696 in Google ADK for Python landed the same point at Google's reference open-source agent framework, where Pillar Security pulled long-lived PATs and GCP service account keys out of the ADK repo's own CI runner through a chain that also started with a prompt in a public artifact.

One prompt is enough. The attack surface here is a public-facing support agent — exactly the shape every enterprise is being told to deploy as the entry point for conversational customer service. There was no admin interface involved, no second-stage code execution, no post-exploitation lateral movement in the classical sense. The prompt was the exploit, the agent was the execution, the metadata was the privilege, and the default role was the blast radius. Every piece was in place when the prompt arrived.

The honest reading

AgentCorruption is patched. The pattern is not. AI agent platforms will keep shipping with ambient credentials inside the agent runtime, because that is how the agent reaches the APIs it was deployed to call. Platforms will keep attaching default IAM roles broad enough to be useful on day one, because narrow defaults produce support tickets. And public-facing agents will keep taking attacker-chosen text as input, because that is their job.

Those three facts together mean the defender's observation point can't live at the IAM layer, the microVM layer, or the model-side alignment layer. It has to live at the layer where the agent decides what to actually do with the message it just received — the layer that sees "the support agent just made an HTTP request to 169.254.169.254" and knows that is not what support agents do.

Zenity reported, AWS fixed, the public now knows. The next platform to ship an AI agent runtime should look at this chain carefully before deciding what metadata its agents can reach by default.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.