Back to blog

How do you secure a system that reads your email and code?

Thắng Đoàn
Thắng Đoàn

One runaway AI agent racked up a $50,000 cloud bill, per Help Net Security. Your agent reads email, queries databases, and runs code, and anyone who tricks it inherits all of that access.

Injection becomes breach when agents hold tools

Prompt injection used to mean embarrassing output. Per Analytics Insight, agentic AI turns it into serious attack potential because agents can read files, call tools, run scripts, and act across business systems. The agent follows instructions wherever they come from, including a poisoned email or webpage. An attacker no longer needs your credentials, just your agent's attention.

The supply chain is now the attack path

Mandiant's AI Risk and Resilience report, cited by Help Net Security, warns that a poisoned data source, model dependency, or extension hook can turn a trusted agent into a channel for internal reconnaissance, lateral movement, or escape from a sandbox. The same report noted attacks have moved from direct prompts to indirect injection and AI supply chain compromises, blurring where your security boundary even sits. The things you plug into the agent are now part of the perimeter.

  • Indirect injection arrives via content the agent reads, not prompts you type.
  • Memory and agent-to-agent channels extend the boundary too, per Analytics Insight.

Shrink what the agent can touch

Verify who is calling, cap what each call can do, and watch behavior live. Analytics Insight recommends least privilege, strict tool controls, identity checks, and runtime policies to limit an agent's attack scope. Mandiant adds adaptive identity controls and reorienting the SOC toward real-time behavioral telemetry. Apple, per Traders Union, updated full disk access requirements on Macs for AI agents. The pattern is consistent: verify who is calling, cap what each call can do, and watch behavior live.

Legal risk is arriving faster than tooling

TechTimes reports the AI Agent Accountability Act, introduced October 1, 2026, would make executives criminally liable if their agent can hack, they knew it, and they skipped safeguards. Apple's annual filing, a public risk-disclosure document companies must publish, warns that agents could exceed permissions, resist instructions, or take hard-to-reverse actions, per Traders Union. Security spend is no longer optional hygiene; it is legal defense.

What this does not solve

None of these controls make injection impossible. Indirect injection rides on legitimate content, so filters stay imperfect. Mandiant's guidance assumes you can observe agent behavior, which many current setups cannot. And as Forbes Technology Council panelists note, many organizations lack the ability to even recognize these AI-specific exposures. Human oversight remains the backstop, and it does not scale.

If your main agent were tricked by one email tomorrow, what exactly could it reach?
Who in your org can see, in real time, what your AI agents are doing?

List every tool, file path, and API your most-used agent can touch, then cut one permission it does not need.

Sources

Analytics Insight: Agentic AI Security

Traders Union: Apple flags AI agent risks

Help Net Security: Mandiant AI risk report

Forbes: Hidden AI Security Risks

TechTimes: AI Agent Accountability Act

Share:

Recommended for you

Enjoyed this article?

Subscribe for new articles. No spam. Unsubscribe anytime.

By subscribing you agree to receive the newsletter. No spam, and you can unsubscribe anytime.