TAU-HOME.COM
LOADING

Verno Labs Dissects Two Types of Prompt Injection in AI Agents and Cuts Hijacking from 84% to Under 2%

Verno Labs categorizes prompt injection attacks on AI agents into tool hijack and argument hijack, revealing distinct internal mechanisms and demonstrating how

tau · November 21, 2026

#AISecurity #PromptInjection #VernoLabs #LLM #AgentSecurity

Verno Labs Dissects Two Types of Prompt Injection in AI Agents and Cuts Hijacking from 84% to Under 2%

In an official research release, Verno Labs has examined the internal mechanisms of prompt injection attacks against artificial intelligence (AI) agents and introduced a combined defense technique that effectively neutralizes them.

Conceptual diagram of AI agent prompt injection defense mechanisms

Image source: https://vernolabs.ai/research/wrong-tool-or-wrong-value-inside-two-kinds-of-prompt-injection

Image Credit: Verno Labs

AI agents frequently process external data from unverified sources such as emails, web pages, and database records. When an attacker hides malicious instructions inside these inputs, the agent can be easily derailed. Verno Labs' research reveals that these attacks are not a monolithic phenomenon; instead, they operate through two distinct internal pathways within the model. Evaluated across open-weight models including Qwen3-8B, Gemma-4-12B-it, Ministral-3-8B, and Llama-3.1-8B-Instruct, the proposed defense framework cuts the average hijack success rate from 84% to under 2%.

The Internal Mechanics of Tool Hijack Versus Argument Hijack

The research team classifies prompt injection attacks into two primary categories based on how they subvert agent behavior:

  • Tool Hijack: The injection forces the model to execute an entirely different tool instead of the one requested (for example, triggering a notification tool instead of paying an invoice).
  • Argument Hijack: The tool selection remains correct, but the parameter values passed to the tool are manipulated in favor of the attacker (such as swapping the payee's bank account ID).

Detailed internal analysis showed that these two attack types rely on fundamentally different neural circuits:

  • Tool Hijack (Global State): The injected text pushes the model into an instruction-following mode represented as a single directional shift in the model's activations.
  • Argument Hijack (Direct Copying): A small set of attention heads—the exact same ones the model normally uses to copy legitimate values out of clean records—pulls the attacker's string straight into the tool call.

Defending via Activation Steering and Field-Level Head Masking

Because the two attack vectors operate differently, traditional single-point defenses typically solve only half the problem. The researchers combined two complementary methods to address both mechanisms:

  • Activation Steering: By measuring the average difference in activations during a tool hijack and subtracting that direction while the model reads the record, the defense neutralizes the global state push that forces instruction-following.
  • Field-Level Head Masking: The agent harness (the application code running the model) uses its JSON parser to identify the declared field being read and hides every other line from the model's 32 primary copy heads. Without access to the untrusted fields, the copy heads naturally fall back to legitimate values.

When combined, these two defenses reduce every attack type to 5% or less (averaging under 2% across models) while preserving normal execution accuracy at 98–100%.

Practical Implications and Limitations

This research demonstrates that agent security must move beyond superficial detection filters and address information flow and internal state control directly at the model and harness level. Because field-level scoping relies on structural metadata rather than heuristic classifiers, it remains robust even when legitimate and injected text are separated by over a thousand tokens.

However, these findings were evaluated on specific open-weight models ranging from 8B to 12B parameters. Further validation is required to test their scalability and transferability across larger proprietary models and complex multi-agent production pipelines.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285