Obfuscated Activations Bypass LLM Latent-Space Defenses

Luke Bailey, Alex Serrano, Abhay Sheshadri, Mikhail Seleznyov, Jordan Taylor, Erik Jenner, Jacob Hilton, Stephen Casper, Carlos Guestrin, Scott Emmons

Abstract: Recent latent-space monitoring techniques have shown promise as defenses against LLM attacks. These defenses act as scanners that seek to detect harmful activations before they lead to undesirable actions. This prompts the question: Can models execute harmful behavior via inconspicuous latent states? Here, we study such obfuscated activations. We show that state-of-the-art latent-space defenses—including sparse autoencoders, representation probing, and latent OOD detection—are all vulnerable to obfuscated activations. For example, against probes trained to classify harmfulness, our attacks can often reduce recall from 100% to 0% while retaining a 90% jailbreaking rate. However, obfuscation has limits: we find that on a complex task (writing SQL code), obfuscation reduces model performance. Together, our results demonstrate that neural activations are highly malleable: we can reshape activation patterns in a variety of ways, often while preserving a network's behavior. This poses a fundamental challenge to latent-space defenses.

Introduction

As Large Language Models (LLMs) are increasingly integrated into critical systems, defending against "jailbreaks"—prompts designed to bypass safety filters—has become a top priority. Traditional defenses typically filter inputs or outputs, but these methods are often brittle. A newer, more sophisticated class of defenses looks "under the hood" at the model’s internal latent activations. The hope is that because activations are causally upstream of the output, they might provide a more robust, "root-level" detection of harmfulness.

This research investigates a fundamental vulnerability in this approach: the malleability of internal representations. We explore whether an attacker can force a model to behave maliciously while simultaneously manipulating its internal "thoughts" to appear perfectly benign to a monitor.

An illustration of how obfuscation attacks bypass defenses compared to traditional attacks. While traditional attacks trigger harmfulness monitors, obfuscation attacks mask internal activations to appear safe.

The Obfuscation Attack Framework

The authors introduce Obfuscation Attacks, which utilize a dual-objective optimization strategy. Instead of solely trying to elicit a harmful response (the behavioral loss), the attacker also minimizes an "obfuscation loss"—essentially the score assigned by the safety monitor. This forces the model to find a path to the harmful output that does not traverse the regions of latent space flagged by the defense.

The study explores several attack vectors:

  • Embedding Suffixes: Adding a sequence of continuous vectors to the input (white-box).
  • Hard Prompt Suffixes: Finding specific strings of text that trigger the behavior (gray-box).
  • Model Finetuning/Data Poisoning: Training the model itself to have "stealthy" backdoors that bypass internal monitors.

These attacks were tested against a wide array of state-of-the-art defenses, including supervised linear and MLP probes, Out-of-Distribution (OOD) detectors (such as Mahalanobis distance and VAEs), and monitors based on Sparse Autoencoders (SAEs).

Key Insights and Results

The Malleability of Representations

The most striking takeaway is that neural representations are highly malleable. There isn't just one "harmfulness direction" in latent space; models can represent the same concept or execute the same behavior using a vast array of different activation patterns. To demonstrate this level of control, the authors optimized an attack to make the model's activation heatmap form a scannable QR code while the model was simultaneously outputting harmful instructions. This proves that an attacker can shape internal states into almost arbitrary patterns without breaking the model's functional output.

Heatmaps showing a harmfulness monitor scan. The bottom-right shows activations forced into a scannable QR code pattern during a harmful generation, successfully fooling the monitor.

Total Bypass of Defenses

Against probes trained specifically to classify harmfulness, obfuscation attacks reduced detection recall from 100% to 0% while retaining a 90% jailbreaking rate. Even Sparse Autoencoders, currently considered a gold standard for mechanistic interpretability, were easily fooled. Attackers could generate complex outputs, like SQL code, without triggering SAE features specifically hand-picked to detect those tasks.

The Complexity Trade-off

Obfuscation is not free. While it is relatively easy to hide the activations for "simple" harmful strings, hiding the activations required for complex tasks (like writing valid SQL) leads to a noticeable drop in model performance. This suggests that some behaviors are more "entangled" with specific activation patterns than others, creating a Pareto frontier between stealth and reasoning capability.

Conclusion and Future Work

The vulnerability of latent-space defenses to obfuscation suggests that internal representations are not the immutable "ground truth" of model intent. The authors attempted "Adversarial Training" for the monitors—retraining the defense on obfuscated activations—but found that even after 70 rounds, attackers could still find new, undetected ways to be harmful.

Future research will need to focus on improving optimization for discrete (text-based) obfuscation and developing more robust defense mechanisms that can withstand the inherent malleability of neural network activations.

Summary updated

Abstract

Recent latent-space monitoring techniques have shown promise as defenses against LLM attacks. These defenses act as scanners that seek to detect harmful activations before they lead to undesirable actions. This prompts the question: Can models execute harmful behavior via inconspicuous latent states? Here, we study such obfuscated activations. We show that state-of-the-art latent-space defenses -- including sparse autoencoders, representation probing, and latent OOD detection -- are all vulnerable to obfuscated activations. For example, against probes trained to classify harmfulness, our attacks can often reduce recall from 100% to 0% while retaining a 90% jailbreaking rate. However, obfuscation has limits: we find that on a complex task (writing SQL code), obfuscation reduces model performance. Together, our results demonstrate that neural activations are highly malleable: we can reshape activation patterns in a variety of ways, often while preserving a network's behavior. This poses a fundamental challenge to latent-space defenses.
BibTeX
@inproceedings{bailey2026obfuscated,
  author = {Luke Bailey and Alex Serrano and Abhay Sheshadri and Mikhail Seleznyov and Jordan Taylor and Erik Jenner and Jacob Hilton and Stephen Casper and Carlos Guestrin and Scott Emmons},
  title = {{Obfuscated Activations Bypass LLM Latent-Space Defenses}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=ktGmDGoWnB},
}
Figure