lamp

adesso Blog

Interview with Barno Kaharova | Part 1 of 2

An ordinary email or document can become an entry point for an attack on a company’s AI system. AI security expert Barno Kaharova explains how malicious instructions can travel between agents. In part two: what traditional security testing misses, and where organisations should start.

Imagine an AI assistant connected to your company’s email, internal documents and business applications. An employee asks it to summarise their inbox. Among the messages is an ordinary-looking email. Hidden in the text is an instruction telling the assistant to retrieve confidential information and send it elsewhere.

The employee has done nothing unusual. There is no suspicious link to click. The attacker is targeting the AI system’s ability to interpret natural language and turn it into actions.

“Before, when we talked about phishing, there was some link you should click. Now it is just a nice email, maybe from your co-worker,” says Barno Kaharova, Senior Consultant and AI Security Expert at adesso SE.

Barno has worked on AI security for several years, including university research and professional projects. Her experience includes testing AI systems and work connected to a test-plan framework for image-based AI systems in agriculture for Germany’s Federal Office for Information Security (BSI).
(Source: BSI publication)

An email that tells your AI agent what to do

The attack in the opening is an example of indirect prompt injection. An attacker places instructions inside material that an AI system is expected to read, such as an email, webpage or retrieved document. The goal is to make the system treat those instructions as commands rather than untrusted content.

One documented example is EchoLeak, a vulnerability disclosed in June 2025 affecting Microsoft 365 Copilot. Researchers showed how malicious instructions embedded in an email could lead an AI assistant to expose sensitive information without the employee clicking a link or approving the action. Microsoft fixed the specific vulnerability before public disclosure.
(Sources: BleepingComputer, Cato Networks, research paper on EchoLeak).

Barno sees EchoLeak as a warning about the wider architecture of AI systems. Fixing an individual vulnerability does not remove the risk when an assistant processes untrusted external content alongside sensitive internal data. “The large language models used in all these agents and copilots are inherently weak against this, because they work with natural language, as the name suggests.”

A filter for suspicious words is unlikely to be enough. “You would have to filter your whole dictionary, because there will always be synonyms.”

How one instruction can travel between agents

Memory poisoning becomes particularly dangerous when multiple AI agents share a knowledge base, persistent memory or one another’s outputs. The risk depends on the system’s architecture and the safeguards around retrieved information.

Barno offers a concrete scenario. An attacker hides a malicious instruction in a requirements document telling an AI agent to disable authentication checks. The first agent reads the document and unknowingly includes that instruction in a summary saved to a shared knowledge base. Later, a second agent retrieves the summary, treats the manipulated requirement as legitimate, and generates code without the authentication checks.

If the code passes review and reaches production, unauthorised users could gain access to customer accounts or sensitive data. The attacker has never needed to modify the source code directly. The instruction has travelled through the agents’ shared information.

“It can be just one sentence in a huge document,” Barno says. The scenario illustrates why the source and trustworthiness of information matter just as much as the instructions an agent receives directly.

That is how one sentence can travel. But what happens when the agent that reads it can also act on it? In part two, Barno explains what an agent should be allowed to do, why traditional security testing is not enough and who should own the risk.

Join us on 25 November

adesso Finland hosts Steps for AI Security & Governance, an executive breakfast session for technology, security and data leaders, on Wednesday 25 November 2026 at Porkkalankatu 5, Helsinki. We start at 08:00 with a delicious breakfast, and the programme begins at 08:30.

Register here: Registration link


Picture Barno Kaharova

Author Barno Kaharova

Barno Kaharova is a Senior Consultant and AI Security Expert at adesso SE in Frankfurt. She works on both offensive and defensive AI security, from data pipelines and models to LLM-powered applications and autonomous agents, and delivers hands-on trainings on AI red teaming, prompt injection, RAG and vector database security, and the OWASP Top 10 for LLM and agentic AI.

Picture Annette Kauppinen

Author Annette Kauppinen

Annette Kauppinen is a Marketing Consultant at adesso Finland. With over 20 years of experience in digital transformation and data-driven marketing, she excels in boosting brand equity and leading cross-functional teams. Annette has successfully driven market expansion across various industries, with sustainability as a core focus in her initiatives.