BSI publishes countermeasures against evasion attacks on LLMs
The German Federal Office for Information Security (BSI) has released “Evasion Attacks on LLMs – Countermeasures in Practice.” The paper addresses a rising threat to AI systems based on LLMs: evasion attacks.
Unlike data poisoning, these attacks happen during operation. Attackers use indirect prompt injections, jailbreaks, or adversarial prompts to bypass or manipulate a model’s intended behavior and safeguards.
BSI recommends practical countermeasures, including:
1) Secure and well-defined system prompts
2) Filtering harmful third-party content
3) Explicit user confirmation before executing sensitive actions
The attached checklist and real-world examples help integrate these measures into existing systems. The guidance targets developers, IT leaders, and public institutions using pre-trained LLMs such as OpenAI’s GPT.