Attackers are increasingly using AI agents to automate all phases of cyberattacks, prompting the security industry and enterprises to find new network defense approaches. One technique that shows promise is to intentionally plant decoy files with prompts that trigger the content safety guardrails built into LLMs with the goal of crashing rogue agentic workflows.
Using decoy resources as tripwires that alert defenders about potential unauthorized access is not a new idea in cybersecurity. These are known as canaries — after the canary in the coal mine early warning system — and can be fake documents, AWS access keys, database dumps, DNS records, and even URLs that would not be queried by legitimate processes, but would be attractive targets for attackers. What’s new in the approach devised and tested by security firm Tracebit is to use these decoy resources not merely to trigger alerts, but to actually stop AI agents, buying defenders more time. Dubbed “context bombing,” the technique takes advantage of the fact that LLMs are inherently vulnerable to prompt injection — acting on instructions they might encounter inside the data they process.
Enterprises that build their own AI agents have to worry about malicious prompts placed by attackers on web pages, emails, documents, code comments, and in other third-party resources those agents might access. With no defenses in place, companies risk their own agents being hijacked and used against them to perform unauthorized actions. But the malicious AI agents used by attackers have the same vulnerability.
“We call the defensive version a context bomb: a short piece of text designed to trigger a model’s safety guardrails, planted directly in the attacker’s path — a decoy secret, environment variable, or DNS record,” Sam Cox, Tracebit’s CTO, said in a blog post. “An AI agent that reads it will frequently refuse to continue. You still get the alert; and the attack now halts early instead of merely being detected.”
Tracebit’s tests performed in an AWS test environment cut the success rate normally achieved by AI models across multiple attacks paths by up to 90% by simply adding instructions that would trigger their content safety guardrails. LLMs are trained to refuse answering questions about creating biological, chemical, or nuclear weapons, or, in the case of Chinese open-weight models, sensitive political topics such as the Tiananmen Square massacre.
Furthermore, once one of these banned instructions make it into the context window, the whole session is poisoned, and the model will often refuse to continue without manual intervention. The researchers ran many tests to find short strings that would trigger model safety guardrails reliably, but it’s worth noting that the identified strings were different between the tested models: Claude Opus 4.8, Gemini 3.1 Pro, GLM 5.2, DeepSeek V4 Pro, and Kimi K2.6.
During baseline tests the AI agents managed on average to obtain full account admin in 54% of the 154 attack runs and full compromise (admin + persistence) in 36% of tests. With the context bombs in place, their success rate dropped to 5% for admin access and 1% for full compromise. Also, in 91% of baseline attack runs, the agents managed to complete at least one of ten possible attacks paths, but their average success rate dropped to 15% with the context bombs.
The models from Western AI labs — Opus and Gemini — proved the most capable at reaching full admin access, with 93% and 70% success rates, but were also the most impacted by the context bombs with both their success rates dropping to 0%. This shows that the content safety guardrails are much stronger in these models compared to the Chinese ones that were tested.
The Tracebit researchers had already been experimenting with AI canaries for alert purposes. A paper published in May showed that canaries alerted defenders in 95.9% of 162 successful attack runs performed with AI agents using 10 different AI models. The alerts were triggered 8 minutes before the agents performed critical actions, while the runs required 14 minutes on average to reach admin access.
During those tests the researchers observed that if they instructed the models that the network might include deception, their success rate for full admin access and persistence dropped from 20% to 3% on average.
Then in June, researchers from security firm Socket.dev came across a software supply chain attack involving malicious PyPI wheels that contained prompt injections with content designed to trigger safety guardrails inside AI models. The malicious prompts were included in code comments at the start of the files and the purpose was likely to evade detection from LLM-powered security scanners.
“This header appears designed for AI-mediated analysis, not for Node, Bun, or Python,” the Socket.dev researchers said at the time. “It attempts to derail scanners or analyst copilots that feed the beginning of a file to a language model without clearly isolating the content as untrusted data. In weak pipelines, this can cause refusal behavior, prompt confusion, context pollution, or premature classification before the scanner reaches the actual malware.”
The Tracebit researchers then had the idea to flip the script and trigger such safety refusals through their canary technique against malicious AI agents. So not only does the agent run stop, but an alert is also triggered because the canary was accessed.
There is a risk that the canaries could also be discovered and accessed accidentally by legitimate LLM-powered tools used by engineering or security teams. But at the same time this means organizations could potentially deploy such canaries in sensitive places to stop their own AI agents that might become hijacked or go off the rails on their own.
There are many reports online where AI models performed destructive or unauthorized actions like deleting databases and folders or elevating their privileges because they got struck in failure loops and explored creative ways to complete their tasks. This is even more common with autonomous AI agents that are tasked to reach a goal without human intervention or supervision.
“The speed of autonomous AI attacks is why deception is climbing the priority list for security programs,” Tracebit’s Cox said. “When the attack chain takes minutes rather than days, every minute of response time you can claw back matters — and a control that stops the attacker outright, rather than just reporting them, changes the economics significantly.”