Skip to content
Security & Trust Talking point

HPE engineer: asking AI agents to ignore hidden instructions fails

An HPE engineering leader says agents that read tickets and email need scoped access, sandboxes and audit trails.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
HPE engineer: asking AI agents to ignore hidden instructions fails
In brief
  • An HPE engineering leader said that asking agents not to follow planted instructions does not work, because the agent bypasses the request.
  • She named scoped short-lived credentials, sandboxes and audit as the fix, and said she has not seen a compliance standard for any of it.

Sadeepa Wijesekara runs engineering for HPE's Private Cloud AI product. On The TWIML AI Podcast she argued that a polite line in the prompt no longer protects anyone. Agents that read tickets, emails and calendar invites, she said, need hard limits on what they can do. She also said she has not seen a compliance standard that says how to set those limits.

What was said

Wijesekara described the attack plainly. Agents run on their own and read content from many places. She said: "Somebody could add some poison information into a ticket or email or a calendar invite." The agent may then take that text for a command and carry it out.

She called the early defence a plea in the prompt, asking the agent not to do this. It does not work, she said, because the agent bypasses it.

Her answer is several layers of control. First, short-lived access tokens. Each token covers one task, and a separate authority grants it. The agent never holds standing credentials. Second, a secure sandbox, where network policy decides whether the agent can reach the internet or other internal systems. Third, confidential computing, which she said keeps even an administrator from seeing prompts in memory. Fourth, logging and auditing everything, with someone watching for an agent that departs from its usual pattern.

She said HPE uses seven layers of security. It sells to federal and defense customers, she added, so its bar is high.

Why it matters

Our reading: the useful question for a buyer moves from how an agent is told to behave to what it is able to touch. Suppose a poisoned ticket reaches an agent holding only a one-task token. The damage stops at that token. An agent with broad standing access has no such stop.

That gives managers a short checklist. Who issues the agent's credentials? Where can it send data? Who would notice if its behaviour changed? A vendor whose only answer is a rule in the prompt has not answered.

The other side

This is one vendor describing its own platform. Wijesekara gave no results showing how well the layers stop attacks. Her point that many public cloud agent runtimes lack confidential computing was a claim in conversation, not a tested comparison.

On standards she was careful. Asked whether any compliance standard says what the right approach is, she said: "I think that I haven't seen it." She invited listeners to point her to one. She also said she was sure people are working on agent security standards. Her own foundation work, she noted, leans toward token economics rather than security. So the gap she describes is her view from one seat, not a survey of the field.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from