Cybersecurity professionals now have a new tool called context bombs. This is according to a study by researchers at Tracebit.

Researchers found that they can confuse AI hacking agents with special prompts. This technique is called context bombing, and it stops the AI attack.

The researchers tested context bombing on five leading AI models. They found that using one context bomb reduced the AI hacking agents' success rate by 90 percent.

The most successful AI hacking agent could access accounts 93 percent of the time without a context bomb. But with a context bomb, it failed every time.

A context bomb is a special kind of prompt. For example, researchers used sensitive topics to stop AI agents on Chinese models. They mentioned Tank Man and Tiananmen Square, which are censored in China.

Context bombs with sensitive topics worked well against Western AI models like Opus and Gemini.

This is the first time defenders have used this technique against attackers.

Attackers have used prompt injections to take over AI assistants. But now, researchers can use them for good as context bombs.

You can find the full study on Tracebit's website.

Topics: Artificial Intelligence, Cybersecurity