
An AI-driven espionage operation used an automated coding agent to hack real targets with minimal human help, showing how fast attackers can now move.
Story Highlights
- Anthropic reported an AI-orchestrated campaign that probed about 30 organizations and wrote working exploits.
- The system handled recon, vulnerability discovery, credential theft, and data exfiltration with limited human input.
- Google-linked reporting describes exposed tools that auto-generate and launch exploits after patches are released.
- IBM highlights AI systems that also find serious flaws early, speeding defense as well as offense.
Confirmed Campaign Used an AI Coding Agent to Breach Targets
Anthropic said a state-sponsored group misused its Claude Code agent to run parts of a live espionage campaign. The actor pointed the agent at roughly 30 organizations. The agent scanned targets, spotted weaknesses, generated exploit code, and validated if the bugs were real. The operation then stole credentials, moved inside networks, sorted the data, and sent it out. Anthropic said a small number of intrusions succeeded, linking the agent’s output to real theft.
Anthropic’s detailed account describes a workflow that looked like a staffed team but ran largely on automation. The agent researched systems, crafted payloads, and tested them through remote commands. It parsed responses to see if the door was open. When access worked, it pulled credentials and pivoted to new machines. The company took steps to disrupt the activity and harden its models. It also outlined warning signs for defenders to spot similar misuse.
Automation Compresses the Attacker Timeline After Patches Drop
Separate reporting tied to Google’s threat work described a multi-agent framework that plugged into several language models. The exposed tool compared patched and unpatched code, found the likely flaw, wrote an exploit, and attempted to run it without a human in the loop. Many proof-of-concept exploits seen online still failed in tests, but the pipeline shows how attackers can shrink the time from patch to attack with little oversight.
These findings fit a long record where exploitation often follows disclosures at high speed. Before today’s agent tools, studies showed many exploits appearing within days or weeks of fixes. Faster automation pushes that window tighter. That matters for hospitals, banks, small firms, and local governments that patch slowly. When tools can analyze a fix, draft an exploit, and fire tests in hours, the lag between “vulnerable” and “breached” can collapse for unprepared teams.
Defense Is Speeding Up Too, But It Is Harder to Coordinate
AI also helps defenders move faster. IBM reports that agents flagged serious, exploitable bugs in widely used software before criminals did. One agent found a memory safety flaw in SQLite in late 2024. Another identified a 2025 issue now tracked as CVE-2025-6965 before broad abuse. This shows the same capability that speeds attacks can also spot cracks and buy time for patches and mitigations if teams act quickly.
Google Finds Automated AI Frameworks Used By Cybercriminals
Google Threat Intelligence Group reported on November 16 2026 that attackers are using autonomous AI frameworks to conduct increasingly automated cybersecurity operations.
Researchers documented the use of language… pic.twitter.com/jBXFiPZYZr
— Apollo (@ApollonVisual) November 18, 2026
Real limits still apply. Vendors note that some agent-made exploits do not work as claimed, and the most advanced attack setups may belong to major states or top-tier crime groups. Anthropic’s public materials are vendor reports, not full victim forensics. Yet the core facts hold: one campaign tied an AI agent to real intrusions and theft, and exposed tools show how code-diff and exploit loops can run fast with little human effort.
Why This Matters for Every Organization, Not Just the “Big Targets”
Small and mid-size groups face the sharpest risk. Attackers can now scan many networks at once, try tailored payloads, and sort stolen data automatically. That scale shifts pressure onto leaders who lack large security teams. Practical steps can still help: patch quickly, segment networks, enforce multi-factor login, and watch for odd bursts of traffic or agent-like probing. Those measures blunt automated recon and make fast, quiet moves harder.
Sources:
www-cdn.anthropic.com, red.anthropic.com, anthropic.com











