News · AI

Imperial and Fraunhofer study cuts LLM prompt injection attack success from 97% to 10%

European researchers developed Heads of Security, using learned per-head attention gates to enforce instruction-data separation in frontier language models.

10%

attack success rate against Heads of Security on RL-Hammer benchmark

Grade 2 · Published method

Source: arXiv

What changed
Researchers at Imperial College London and Fraunhofer HHI introduced Heads of Security, cutting black-box prompt injection attack success rates from 97% to 10%.
Who it counts
The defense was evaluated across Llama-3.3-70B-Instruct and GPT-OSS-20B against black-box and adaptive white-box adversarial attacks.
What it does not tell you
The benchmark measures static and synthetic prompt injection datasets and does not evaluate multi-turn social engineering or multi-modal attacks.

The full story Check the source

Researchers from Imperial College London and Fraunhofer HHI developed an architectural prompt injection defence that reduces black-box attack success from 97% to 10%, according to a study published on arXiv [1].

The method, named Heads of Security, learns per-head attention gates that redistribute attention weights between trusted system instructions and untrusted data tokens [1]. In evaluations on GPT-OSS-20B, the RL-Hammer attack achieved a 97% attack success rate against the baseline model and 85% against fine-tuned defence SecAlign, but dropped to 10% when protected by Heads of Security gates [1]. Adaptive white-box attack Claudini required 3 times more compute FLOPs to achieve the same breach rate against the gated architecture [1].

What it measures

The study measures empirical attack success rates and benign task utility across Llama-3.3-70B-Instruct and GPT-OSS-20B using standardized prompt injection benchmarks [1]. Utility retention was measured across multiple-choice reasoning, standard few-shot prompts, and agentic tool-use environments [1].

What it does not tell you

The evaluation does not measure defence performance against multimodal inputs containing visual steganography or adversarial audio payloads [1]. The paper does not report deployment latency overhead or memory footprint implications when serving gated attention layers on commercial inference engines [1].

Sources and their grades

  1. arXiv, Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness. Read 28 October 2026. Grade 2 · Published method arXiv:2610.37657v1 [cs.CR, cs.AI], submitted 2026-10-27 17:58:52 UTC by researchers from Imperial College London and Fraunhofer HHI / TU Berlin. Stage 2 computer science preprint released on arXiv; not peer-reviewed.

A grade says how much weight a source will take, from 1 (a body that can compel an answer and has nothing to sell) to 5 (somebody describing what they sell, or themselves). The standard sets out what each grade may be used for.

Filed under AI research, AI safety, Cybersecurity

Comments

Commenting needs an account, made with nothing but an email address. A first comment is read before it appears.

More news

See all