Germany’s BSI issues guidelines to counter evasion attacks targeting LLMs

Germany’s BSI has published guidance on evasion attacks in large language models, warning that such attacks can subvert model safeguards. The document, titled Evasion Attacks on LLMs – Countermeasures in Practice, is aimed at developers and IT managers in organisations using pre‑trained models such as GPT, as well as other experienced IT users.

It details evasion methods like prompt injection and data manipulation and recommends secure prompts, filtering, Zero Trust, and anomaly monitoring to raise the attack cost and reduce potential damage. The BSI advises combining technical controls—such as filters, sandboxing, and retrieved‑augmented generation with trusted retrieval—with organisational practices, including adversarial testing, governance, and training, as part of a defence‑in‑depth approach.

It emphasises that no single control is sufficient and that layered safeguards and continuous monitoring are essential as organisations adopt LLMs. According to the publication, the aim is to help secure AI systems used by companies and public authorities against evolving evasion threats.