Tracebit, a company operating out of both London and New York, has revolutionized the defense against AI-driven attacks by deploying innovative ‘context bombs.’ These mechanisms are akin to landmines in a military operation, designed to divert and confuse AI agents during their attempts to breach security. When faced with these carefully crafted prompts, the attacking AI inadvertently processes misleading information, undermining its original objectives.
Advanced Tactics to Misdirect AI Agents
Building on previous research, Tracebit’s latest study titled ‘Making Attacking Agents Pay’ introduces sophisticated methods for derailing AI attacks. By incorporating nine new context techniques, attackers are led on a wild goose chase involving false targets and convoluted puzzles. This approach effectively neutralizes attempts by advanced models like GPT 5.6 Cyber, Kimi K3, Qwen 3.8 Max, and DeepSeek V4 Pro. Notably, the strategy also significantly reduced the success rate of GPT-6 Astra by 54%, while practically obliterating Claude Opus 5’s attack efficacy by 91%.
Effective Deceptive Techniques
A standout among these prompted deceptions is a fictitious pen testing report. It masquerades as an authentic analysis, mislabeling real attack paths and highlighting decoy targets. This tactic led AI agents to perceive genuine systems as traps, thereby evading them and decreasing their success by 58%. Some systems even halted attacks prematurely, leaving potential vulnerabilities untouched. This resulted in a dramatic drop in objective completion across systems like Cyber, Kimi, Qwen, and DeepSeek from 39 successful attempts to merely 2.
Combining Strategies for Maximum Impact
Additional context bombs include endless web services, seemingly encrypted password layers, and a deceptive infinite database. In one instance, the GLM-5.3 model spent 55 minutes navigating 825 pages only to realize it was ensnared in a loop. The collective deployment of all nine context bombs proved especially potent against Chinese models from Moonshot, Alibaba, and DeepSeek, completely thwarting their objectives.
Even when these context bombs didn’t fully prevent an attack, they significantly raised the cost by 23 times per successful breach. Although attackers may possess substantial resources, the inflated cost often outweighs the potential benefits, discouraging further attempts.
Tracebit’s co-founder and CEO, Andy Smith, articulates the strategy: “In the face of increasingly deceptive AI behaviors, our best defense lies in counteracting these tactics with our own. By leading them through intricately deceptive pathways, we exhaust their resources and heighten the likelihood of triggering alerts, providing defenders with crucial response time.”
