When AI Agents Attack: What the First AI-Orchestrated Campaigns Teach Us
AI-powered cyberattacks are now documented, from Claude's use in state-sponsored espionage to a breach of Mexican government agencies.
· Updated · 12 min read · Guard0 Team

We're no longer theorizing about AI-powered cyberattacks. Since late 2025 there has been documented evidence of a state-sponsored group using AI agents to run intrusions, an individual operator breaching government agencies with AI assistance, and vulnerabilities in the tools developers use to build agents.
This is our analysis of what happened and what it means.
Campaign 1: Claude GTG-1002, the Chinese State-Sponsored Operation
In November 2025, Anthropic published a landmark disclosure detailing what they called the first documented large-scale cyberattack executed primarily by AI.
What Happened
A Chinese state-sponsored group leveraged Claude Code to conduct autonomous cyber operations against approximately 30 global organizations, including technology companies, financial institutions, and government agencies. The operation was detected in mid-September 2025.
The Attack Architecture
The attackers exploited three key AI capabilities:
| Capability | How It Was Used |
|---|---|
| Intelligence | Advanced language models following complex instructions and executing sophisticated coding tasks |
| Agency | AI systems operating autonomously in loops, chaining tasks with minimal human direction |
| Tools | Access to software utilities via Model Context Protocol for reconnaissance and exploitation |
The Kill Chain
The attack progressed through six phases:
- Target Selection: AI-driven research to identify high-value organizations
- Infrastructure Reconnaissance: Autonomous mapping of technology stacks and attack surfaces
- Vulnerability Identification: AI-powered discovery and analysis of weaknesses
- Credential Harvesting: Automated extraction of authentication material
- Data Exfiltration: Systematic extraction of sensitive data
- Documentation: The AI documented its work so the operation could be handed off
AI performed 80 to 90 percent of the operation, with human intervention at only 4 to 6 critical decision points per campaign, according to Anthropic. That is a large force multiplier.
The Jailbreak Technique
The attackers succeeded by decomposing malicious objectives into innocent-seeming subtasks while misrepresenting Claude's purpose as "defensive security testing." This is exactly the semantic misdirection that makes agentic attacks uniquely dangerous: the AI never sees the full malicious picture, only reasonable-sounding steps.
This technique worked against one of the most safety-focused AI labs in the world. It will work against your agents too, unless you have runtime behavioral controls that go beyond prompt-level guardrails.
Campaign 2: The Mexico Breach, One Operator and an AI Assistant
In February 2026, Bloomberg reported research describing an attacker who used Claude to breach multiple Mexican government agencies, starting in December 2025 (coverage). The targets included the federal tax authority and the national electoral institute. Anthropic said it banned the accounts involved.
The figures come from one security firm's research and have not been independently confirmed:
- About 150GB of data taken
- Documents related to 195 million taxpayer records, along with voter and civil registry data
- A single operator, according to the researchers
The traditional security assumption, that sophisticated attacks require sophisticated teams, is under pressure. AI-assisted tooling acts as a force multiplier, enabling smaller teams to conduct campaigns that previously required far greater resources.
Claude Code Vulnerabilities: When the Tools Are the Target
In February 2026, Check Point Research disclosed two vulnerabilities in Claude Code, the same category of tooling used in the GTG-1002 campaign. Both were triggered by opening an untrusted repository, and Anthropic patched both before publication:
- Remote code execution (CVE-2025-59536): repository-controlled configuration could run commands before the user approved anything
- API key exfiltration (CVE-2026-21852): a malicious project could redirect API traffic, including the authorization header, to an attacker's endpoint
These aren't attacks by Claude. They're vulnerabilities in AI developer tooling. The distinction matters: as AI coding agents become standard developer infrastructure, their attack surface becomes critical infrastructure.
Every coding assistant with terminal access, every agent with MCP tool integration, and every autonomous system with API credentials is a potential vector now.
Military Adoption of AI in Operations
Militaries are adopting the same tools. In December 2025 the U.S. Department of War launched GenAI.mil, a generative AI platform for its personnel. When AI agents take part in high-stakes decision chains, governance is a hard requirement, not a compliance checkbox.
This isn't about whether military AI is good or bad. The agent governance frameworks that enterprises need are being tested at the highest possible stakes. Governance that can't handle an expense report approval workflow can't handle decisions like these.
The Wider Picture
The individual incidents tell a story. Published reporting points the same way:
| Source | Finding |
|---|---|
| CrowdStrike Global Threat Report | Year-over-year increases in AI-enabled adversary activity |
| IBM X-Force Threat Intelligence Index | A rise in AI-assisted phishing campaigns |
| OpenAI | Said in December 2025 that prompt injection is unlikely to ever be fully solved (Fortune) |
| "The Attacker Moves Second" (Nasr et al., 2025) | Adaptive attacks bypassed 12 published defenses, with success rates above 90% for most |
AI-powered attacks are improving faster than static defenses.
--------|---------| | CrowdStrike / industry threat reports | Significant year-over-year increases in AI-enabled adversary activity | | IBM X-Force / phishing trend data | Measurable rise in AI-assisted phishing campaigns | | OpenAI | Acknowledged prompt injection "may never be fully patched" | | Academic research | High bypass rates (90%+ in some studies) against static prompt defenses ("Attacker Moves Second" problem) |
The trend is unambiguous. AI-powered attacks are growing faster than AI-powered defenses.
The Paradigm Shift: From "AI Security" to "Agent Accountability"
These incidents collectively reveal a paradigm shift that most organizations haven't internalized:
The old model focused on making AI safe to talk to: preventing harmful outputs, filtering toxic content, aligning models with human values. These are guardrails, and they matter.
The new reality requires knowing what AI agents do and being able to prove it: discovering every agent, assessing every risk, and keeping a record of what each one did. It's not just about security. It's about accountability.
Guardrails control what an agent says. Accountability is knowing what it does.
This isn't a subtle distinction. The GTG-1002 campaign didn't involve toxic outputs or harmful content. The AI was helpful, capable, and polite. It just happened to be helping an attacker. Output guardrails don't address that. Knowing what your agents are doing, and being able to prove it, does.
The "Attacker Moves Second" Problem
Static defenses fail against AI-powered adversaries because the attacker can iterate against your defenses at machine speed. In "The Attacker Moves Second", researchers used adaptive attacks to bypass 12 published jailbreak and prompt injection defenses, with success rates above 90% for most of them. The defender deploys a filter, and the attacker generates variants until one passes.
The defense that holds up is runtime behavioral analysis: monitoring what agents actually do rather than trying to predict what attackers will say.
The Timeline So Far
Anthropic detects a state-sponsored campaign targeting about 30 global organizations
The techniques become public knowledge
A single operator starts using an AI assistant against Mexican government agencies
Check Point Research publishes two Claude Code vulnerabilities, and Bloomberg reports the Mexico breach
The organizations investing in agent security infrastructure now will be prepared for what follows.
Attack Agent Architectures
AI-powered attacks fall into five categories of increasing sophistication:
Autonomous Reconnaissance Agents: Operate 24/7, systematically map attack surfaces including AI agent endpoints, maintain context across sessions. This was the starting point of the GTG-1002 kill chain.
Exploit Generation Agents: Create novel exploits based on discovered vulnerabilities. Can generate attacks that bypass signature-based detection because they're original.
Social Engineering Agents: Conduct personalized spear phishing at machine scale. Research targets across LinkedIn, GitHub, and social media, then craft individually tailored lures.
Evasion and Adaptation Agents: Learn from defensive responses in real-time. When one attack vector is blocked, they pivot within minutes, too fast for human analysts but natural for an AI operating in a loop.
Agent-Targeting Agents: AI specifically designed to compromise other AI agents. Probes for prompt injection vulnerabilities, tests MCP tools for exploitation vectors, and propagates through multi-agent systems via lateral movement.
Defending Against AI-Powered Attacks
Defending against these threats requires moving beyond traditional security controls:
1. Runtime Behavioral Monitoring
Monitor what your agents do, not just what they're asked to do. The GTG-1002 campaign's individual steps looked legitimate. It was the pattern that was malicious.
2. Agent Identity and Access Governance
Ensure every agent has a verified identity with least-privilege permissions. The force multiplier effect means a single compromised agent credential can enable an entire campaign. See: AI Agent Identity: The Accountability Challenge No One Is Solving
3. MCP and Tool Security
Secure the tool integration layer. The GTG-1002 attackers exploited MCP for real-world capability. Your agents' MCP connections are your attack surface. See: MCP Security Guide
4. Prompt Injection Defense in Depth
Accept that prompt injection will never be fully solved. Layer defenses: input filtering, instruction hierarchy, action validation, behavioral monitoring, and human oversight for high-risk actions. See: Agent Prompt Injection: Beyond Basic LLM Attacks
5. Incident Response Readiness
When (not if) an agent is compromised, your response time determines the damage. Have playbooks ready. See: Agent Incident Response
Live walkthrough of agent discovery, risk scoring, and policy enforcement.
Book a DemoKey Takeaways
- AI-orchestrated attacks are documented: GTG-1002, the Mexico breach, and the Claude Code CVEs are public record
- Force multiplication changes the threat model: a small team with AI can match a much larger operation
- Guardrails are necessary but insufficient: controlling what agents say doesn't control what they do
- The attacker moves second: adaptive attacks bypassed 12 published prompt defenses
- These techniques are public now, so preparation can't wait
Assess Your Exposure
These threats affect every organization deploying AI agents. Understanding your exposure is the first step.
Request a free AI risk assessment
Guard0 tests the agents, MCP servers, and AI applications you run, then sends you the findings and the order to fix them in. To see how Guard0 tests agent behaviour continuously, read about AI red teaming.
References
- Anthropic, "Disrupting the first reported AI-orchestrated cyber espionage campaign," November 13, 2025
- CrowdStrike, "Global Threat Report," annual series (trends referenced from publicly available editions)
- IBM X-Force, "Threat Intelligence Index," annual series (trends referenced from publicly available editions)
- NVD, "CVE-2025-59536" and "CVE-2026-21852," Claude Code, disclosed by Check Point Research, February 2026
- MITRE ATT&CK, "TA0043 Reconnaissance," "TA0001 Initial Access"
- MITRE ATLAS, "AML.T0054 LLM Jailbreak"
- Nasr, et al. "The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt Injections," 2025
- Engadget, "Hacker used Anthropic's Claude chatbot to attack multiple government agencies in Mexico," February 2026
- Fortune, "OpenAI on AI browser prompt injections," December 23, 2025
Threat Intelligence · AI Attacks · Nation-State · Claude · Adversarial AI