Your Coding Agent Found Your AWS Credentials. Who Answers for What It Does Next?

Your Coding Agent Found Your AWS Credentials. Who Answers for What It Does Next?

BackerLeader ●58 ●371 ●551
calendar_today ago • schedule5 min read

In one early HYPR deployment, an engineer was doing design work with an AI agent on a work laptop. The agent needed details about the AWS environment. The engineer was already signed in through the AWS CLI. So the agent used those credentials and started running commands against AWS. Nobody told it to.

A policy caught it. The gateway inspected the agent's tool call, saw AWS CLI command lines, and blocked the request.

"The commands weren't destructive, but it's easy to see how they could have been," said Bojan Simic, CEO and co-founder of HYPR. "Interestingly, the agent saw nothing wrong with its actions."

That last line sums up the problem. Coding agents are good at finding a way to finish the job. If a credential is lying around, they'll use it. And you can't fix that by asking nicely. "Telling an agent not to do something again is still non-deterministic," Simic said.

Most agents today run on borrowed access: a developer's API key, a shared service account, a CLI session somebody left open. Nothing ties the agent's action back to a person. Security teams can't tell which agent did what, or for whom.

HYPR is going after that gap with AgentPass, which launches today. The idea is simple to state. Every agent gets bound to a named human supervisor. Every request gets checked against identity and policy before it runs. High-risk actions go to that human for a passkey-signed approval. And an authorized user can shut an agent off on its next request.

What an agent's identity looks like

AgentPass routes agent traffic through a single gateway, whether the agent runs on a laptop, in a customer's container on AWS, or on a provider-hosted platform like Microsoft 365 Copilot or AWS Bedrock AgentCore. Three identities can come into play.

First, the agent gets a token of its own, kept in sync with the organization's identity provider. That makes every request traceable to a specific agent. Second, for tools the supervisor is already allowed to use, the agent uses the supervisor's identity. It never gets more access than that person has. Third, when the supervisor delegates a specific action, the agent gets a short-lived session token scoped to that one delegation.

Agents start with zero access. They see only the tools their supervisor's role allows, and they never hold the credentials for those tools. The gateway keeps them and hands out short-lived, single-tool tokens. Because access follows the supervisor's role, a role change changes what the agent can do. Offboarding a person cuts off their agents on the next request.

"There is no agent without an accountable person," Simic said. That includes workload agents that run unattended. Those get a role of their own instead of borrowing someone's access. A policy can also name who approves a type of action, such as a production change, no matter which agent attempts it.

A new agent's first connection is held until the supervisor approves it with a passkey on their phone. That approval expires and has to be renewed.

What a developer notices

On macOS and Windows workstations, a lightweight Endpoint Guard looks for tools like Claude Code, Codex, and Cursor. It uses code signatures, process identity, and network destinations, then reports what it finds under a cryptographically signed device identity. Anything that doesn't match a known tool is dropped on the device and never sent.

For tools a company chooses to manage, the guard configures the tool to send traffic through the gateway. "This is a setting, not a network proxy," Simic said. It also installs the approved MCP tools and blocks the tool from talking directly to its AI provider, so nobody routes around the gateway.

Day to day? "Mostly nothing," Simic said. Developers approve an agent's first connection, get the occasional approval request on their phone, and see a clear message when policy blocks something.

Plain-English policies, compiled rules

Administrators write policies in plain language, such as "stop the agent from writing files outside the project directory." An AI model drafts a precise rule from that. But the model isn't involved at enforcement time. The gateway runs the compiled rule the same way every time, outside the agent's model, so a prompt injection can't talk its way past it.

There are checks before a rule can block anything. The rule is explained back in plain language, with the exact logic open for inspection. Ambiguous requests trigger a clarifying question. Rules that won't compile get rejected. A draft can be replayed against the last seven days of real traffic. And every new rule starts in Observe mode, recording what it would have blocked until a person promotes it.

"AI helps write the rule, and a person decides whether it gets enforced," Simic said.

The approval fatigue problem

Any human-in-the-loop system runs into the same wall. Send enough prompts and people start tapping "approve" without reading.

HYPR's answer starts with volume. "The best defense is asking less often," Simic said. Routine work is handled by allow and block rules and by limiting which tools a role can reach. The company recommends a cap of two to three approvals per person per work day.

When a prompt does show up, it goes to the supervisor's phone, a separate device from the one the agent runs on. It shows which agent is asking and what it wants to do, with a preview. Approving takes a biometric-backed passkey. Unanswered requests expire and are never silently approved. Every decision is signed and logged, so a pattern of reflexive approvals shows up in the audit trail.

What early deployments turned up

HYPR says it typically finds unsanctioned AI activity across roughly 20% of the workforce in early deployments. Affected users often have two or three unsanctioned AI tools or services. At the gateway, the company says it generally finds two or three unsanctioned or unauthenticated agents per day. These are HYPR's own figures, drawn from its gateway and endpoint data. HYPR didn't share how much standing access teams removed.

Where the edges are

The most useful part of Simic's answers was his warning about vendors, his own included.

"Be skeptical of 'we see and control everything your agents do,'" he said. A gateway sees only the traffic that passes through it. An endpoint agent sees only the machines it's installed on. And because coding agents run commands locally, some actions can only be caught in the model's response, not as they run.

His advice is to make vendors show you, not tell you. Point an agent straight at the provider and see if the bypass gets blocked or at least detected. Disable an agent mid-task and check whether its next request fails on every node. Install a tool the vendor didn't plan for and time how long it takes to show up. Ask whether an approval is signed by a specific person on their own device or is just a button click. And ask to see the exact rule that will run, test it on your traffic, and watch it before it enforces.

That's a good checklist for any agent security product, not just this one.

Here's my take. Whoever builds or runs an agent is responsible for what it does, and so is the company that employs them. Most organizations already believe that. What they've lacked is a way to put a name on every agent action and prove it. AgentPass is one attempt to make that accountability enforceable. Run the tests Simic describes and see how well it holds up.

HYPR AgentPass is available now. More details are at hypr.com/agentpass.

Want more #agenticai & #aisecurity deep dives?

Join 5,953+ developers sharing real benchmarks, code breakdowns, and projects. Free forever.

More Posts

Your AI Agent Doesn't Need Your Token: Edera's Ariadne Conill on Object Capabilities

Tom Smithverified - Nov 5

Cisco's Amy Chang: A Model's "Passport" Doesn't Tell You Where It Actually Came From

Tom Smithverified - Aug 27

What Security Teams Want From AI Agents Before They Let Them Act Alone

Tom Smithverified - Nov 18

AAIF Hits 10 Projects. Its CTO on What to Adopt Now and the Security Gaps Left Open

Tom Smithverified - Oct 23

AI Agents Don't Have Identities. That's Everyone's Problem.

Tom Smithverified - Mar 13
chevron_left
21k Points • 980 Badges
289Posts
171Comments
143Connections
LLM Training & Evaluation Specialist with hands-on experience building major AI models. As one of th... Show more

Related Jobs

Commenters (This Week)

1 comment

Contribute meaningful comments to climb the leaderboard and earn badges!