AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

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

ICLR 2026 · Adversarial robustness and security › Backdoor and poisoning · approach: Backdoors in large multimodal models · 40 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add an obfuscation loss to embedding suffix attackprobe-evading jailbreak · Harmfulness probesRecall @ 1% FPR (%): 100% -> 0%Improved
use generation-based instead of instruction-based mean difference probejailbreak detection · Jailbreak detectionRecall @ 1% FPR (%): 39.66 -> 71.94Improved
use MLP instead of mean difference probe (generation-based)probe · Jailbreak detectionRecall @ 1% FPR (%): 71.94 -> 96.29Improved
train obfuscation suffix against OOD detectorinput-based probes · Obfuscation transferRecall @ 1% FPR (%): bypasses input-based probesImproved
use circuit breakers trainingLlama-3-8B-Instruct · Embedding suffix attackJailbreak success rate (%): 95%Hurt
add obfuscated adversarial training (OAT)Llama-3-8B-Instruct · Obfuscation suffix attackJailbreaking rate (%): 95% -> 35%Improved
add obfuscated adversarial training (OAT)Llama-3-8B-Instruct · Obfuscation suffix attackGradient steps to reach zero recall: 256 -> 16,384Improved
add adversarial training of monitors for 70 epochsMLP probe · Harmfulness monitoringRecall on new attackNo clear effect
add obfuscation loss in backdoor training (OOD and probes)backdoored model · Backdoor obfuscationAttack success rate (%): 95.5 -> 94.9No clear effect
add obfuscation loss in backdoor training (OOD and probes)backdoored model · Backdoor obfuscationFalse trigger rate (%): 0.64 -> 0.96No clear effect
raise SAE raw latent probe threshold from 1% to 2% FPR (ablation)SAE raw latent probe · SQL detection with obfuscation suffixRecall (%): 0% -> 100%Improved
increase number of poisoned samples (ablation)backdoor obfuscation · Backdoor with simple triggerobfuscation vs all detectors: around 200 poisoned samplesImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy