A fake secret makes the attacker's AI agent shut itself down
In July 2026 the security company Tracebit published a working paper on what it calls context bombs. The idea is to place a short piece of text as the content of a decoy resource, for instance as the value of a fake secret in AWS Secrets Manager. When an attacking AI agent reads it, it meets content that triggers the model's own guardrails, and the session stops. In the experiments the share of runs where the attacker reached full administrator access fell from 57 to 5 per cent. Ars Technica covered the work on 13 July 2026.
What happens technically
The technique works in two ways at once. The attacker's own model refuses to continue, and the defender is simultaneously alerted that the decoy resource was read, which is how decoys are meant to work. The content has to be tailored to the model it is meant to stop. Tracebit states that content about biological weapons stops the leading Western models, while politically sensitive content in Chinese stops models from Chinese suppliers. The two concrete examples in the paper are anthrax spores and Tank Man from Tiananmen Square in 1989. Tracebit also mixed the sensitive content with ordinary prompt injection devices such as urgency markers and delimiters, and states that which device worked best varied from model to model. The demonstration on its site shows an agent reading a fake secret and receiving an error from its supplier saying the content carries risk, whereupon the session ends.
The rig is a constructed AWS environment of around 300 resources with ten attack paths, where the agents were cast as red teams with initial access through a low privileged key. The figures rest on five models and 152 runs, with a single bomb in a single decoy secret. Full compromise fell from 36 to 1 per cent, the share of runs completing at least one attack path from 91 to 15 per cent, and the number of completed attack paths per run from 1.53 to 0.16. The effect varies sharply between models: Claude Opus 4.8 went from 93 to 0 per cent administrator access and Gemini 3.1 Pro from 70 to 0, while Kimi K2.6 went from 27 to 11. Tracebit notes several caveats itself. The test setup was not identical for every model, OpenAI models were not tested in the rig, and models with their guardrails removed were not examined. The company states plainly that the technique does not stop AI attackers completely, and that a stopped attack is not a closed attack. Note as well that Tracebit sells a product carrying this feature, and that the effect at customers has not yet been measured. The sources differ slightly on the numbers: CSO Online reports 154 runs and 54 per cent, while Tracebit's own figures are 152 and 57.
What this means for you if you use decoy resources in AWS
If you already run decoy resources in your environment, this is a change of content rather than a new tool. You replace the value in a fake secret with something that both alerts you and stops the reader. Our assessment is that this is exactly why the technique is interesting: the cost is low, and what you get is a brake that works against the kind of attacker who otherwise moves faster than you can respond. Note at the same time what the numbers actually say. The strongest attackers fell hardest, while the weakest were least affected. You get the most benefit against the opponent you fear most.
Two side effects need planning before you try this. One is that the bomb can also stop your own legitimate AI agents, and Tracebit's advice is to choose content aimed at suppliers other than the ones you use. The other is that the strings can alarm people who stumble across them in an operational environment, and there Tracebit suggests encoding the text. Remember finally that this is one company's experiment in one rig, with that company's own product at the other end. It is enough to try in a controlled environment, and not enough to build your defence on.
Berigo recommends
- Try the technique in a controlled environment first, and only where you already have decoy resources and alerting that works.
- Choose content aimed at model suppliers other than the ones you use yourselves, so your own agents are not stopped.
- Treat a stopped attack as an alert rather than as a resolution. Investigate how the attacker got that far.
- Encode the text, so that operations staff do not meet offensive content in their own systems without warning.
- Do not swap basic measures for this. Narrow privileges and alerting still decide the outcome.
Security that is understood, governed and works.
Let us help you turn security into an advantage, not a cost. Get in touch for a no-obligation conversation about where your organisation stands and what to prioritise first.
Get in touch