Policy

Criminal Hackers Are Exploiting Gaps Between AI Guardrails to Run Autonomous Attack Campaigns

Threat groups are bypassing model safeguards by mixing older Claude and Deepseek models, routing tasks based on cost, and using social engineering—not novel exploits.

Last verified:

Four previously unreported threat groups ran successful autonomous hacking campaigns between May and August 2026 by exploiting inconsistencies in AI model safeguards, according to threat intelligence firm CrowdStrike. Rather than discovering new vulnerabilities, the attackers leveraged differences in guardrail strength across model versions and orchestrated multi-model attack pipelines to evade restrictions—a pattern that underscores a critical gap in AI security governance as malicious actors adopt agentic AI at pace with defenders.

Model Routing as a Bypass Strategy

The threat groups—operating from China, Spanish-speaking regions, Indonesia, and Russia—deliberately mixed older versions of Claude and open-weight Deepseek models within single attack campaigns, routing different tasks to different systems based on both cost and guardrail weaknesses. According to CrowdStrike’s senior vice president of counter adversary work, Adam Meyers, attackers “leverage different models for different aspects in the operation. Model routing and selection is very important because it’s expensive and a lot of times would be overkill to use newer models.”

This approach inverts the typical narrative about AI security: rather than brute-forcing a single model’s protections, adversaries treat guardrails as a routing problem to be optimized. Meyers notes that “the models don’t all have equivalent guardrails and protections, even from one company,” meaning attackers can shuffle tasks to systems with lower barriers. The Chinese and Spanish-language groups displayed sophisticated orchestration infrastructure; the Indonesian and Russian actors showed less technical maturity but still succeeded in compromising multiple targets—suggesting the tactic’s accessibility across skill levels.

Social Engineering Over Novel Exploits

All four groups circumvented safeguards using variants of established jailbreak methods rather than discovering zero-day vulnerabilities. The attackers framed malicious activity as authorized penetration testing or security evaluation, added forged approval files to agent working directories to simulate legitimacy, and crafted prompts designed to preempt model objections. Meyers characterizes the attacks as exploiting “the speed and the volume that you can use AI for in conducting these operations targeting the bulk of security hygiene issues in a lot of these environments—things that were preventable but weren’t prevented.”

The fact that all four actors left their AI infrastructure openly accessible on web servers allowed CrowdStrike researchers to observe their model selection and operational patterns—revealing that none prioritized frontier models, favoring older versions to manage compute costs. This accessibility, paradoxically, provided rare visibility into active criminal use of AI systems.

Why This Matters

The findings expose a structural weakness in the AI safety approach: guardrails are being treated as patchwork barriers rather than coherent systems. If attackers can achieve operational success by routing around inconsistent protections across model families, then vendor-by-vendor and version-by-version safety improvements may leave coordinated adversaries with ample workarounds. Security teams evaluating AI deployment should assume that model safeguards will be tested by well-resourced threat actors—and that the bar for bypass is lower when multiple models with varying protections are available. Conversely, the report suggests that most active criminal AI use is not yet targeting zero-day model vulnerabilities, but rather scaling attacks on exploitable infrastructure and human security failures—a more tractable, if unglamorous, defense problem.

Frequently Asked Questions

How are attackers bypassing AI safety guardrails?

According to CrowdStrike, threat actors are using social engineering (framing attacks as authorized penetration tests), adding forged approval files to agent working directories, and crafting jailbreak prompts that pre-emptively address model concerns.

Which AI models are being abused in these attacks?

The four threat groups relied primarily on older versions of Anthropic's Claude and open-weight Deepseek models, rather than frontier-generation systems, to reduce costs while maintaining sufficient capability.

What makes these attacks hard to stop if they don't use new exploits?

CrowdStrike's Adam Meyers notes that the attacks exploit existing security hygiene failures and leverage AI's speed and volume to target preventable vulnerabilities at scale—not novel attack surfaces.

Do all AI models have equivalent safeguards?

No. According to Meyers, guardrails differ both across companies and between model versions within the same company, allowing attackers to route sensitive tasks to models with weaker protections.

#ai-safety #cybersecurity #jailbreaking #guardrails #threat-intelligence