The First LLM Agent Cyberattack: How an AI Hacker Exfiltrated a Database in Under an Hour
On May 10, 2026, something quietly changed in the cybersecurity landscape. A server somewhere on the internet, running an open-source Python notebook platform called Marimo, was compromised through a critical vulnerability. What happened next was unlike anything security researchers had seen before — not because of the intrusion itself, but because of who — or rather what — was driving it.
An LLM agent took the wheel.
Sysdig's Threat Research Team (TRT) has published what they believe is the first confirmed in-the-wild attack in which an AI agent autonomously conducted the entire post-exploitation chain — making decisions, adapting to outputs, and pivoting through infrastructure without human direction between steps. From initial access to a fully exfiltrated internal PostgreSQL database, the entire operation took under 60 minutes. The SSH bastion phase alone — the final pivot — completed in under two minutes.
The Attack Chain: Four Pivots, Zero Human Input
The attack unfolded in four distinct phases, each executed by the LLM agent without manual intervention:
Step 1: Initial Access via CVE-2026-39987
The attacker exploited CVE-2026-39987, a critical pre-authentication remote code execution vulnerability in Marimo, an open-source Python notebook platform. A single WebSocket request granted an interactive shell on the compromised host.
Unlike traditional attacks where a human operator would take over at this point, the attackers handed control to an LLM agent — an AI system designed to analyze environments, formulate next steps, and execute commands in real time.
Step 2: AWS Credential Harvesting
The agent immediately began reconnaissance. It extracted two cloud credentials from the compromised host by reading environment variables, configuration files, and metadata endpoints. Sysdig's analysis showed the agent issuing targeted commands rather than running generic enumeration scripts — a hallmark of AI-driven behavior.
Step 3: SSH Key Retrieval from AWS Secrets Manager
With cloud credentials in hand, the agent called the AWS Secrets Manager API to retrieve an SSH private key. Critically, it used a fanned-out egress pool — 12 cloud API calls across 11 distinct IPs in just 22 seconds, leveraging Cloudflare Workers as distributed exit nodes. This distributed approach broke traditional IP-based alerting entirely.
Step 4: Database Exfiltration Through SSH Bastion
The agent opened eight parallel SSH sessions through a bastion server, dumping host configuration and the complete contents of an internal PostgreSQL database. This phase — from the first SSH connection to full exfiltration — took less than two minutes.
How Researchers Knew It Was an AI Agent
Sysdig identified several telltale signs that an LLM agent, not a human, was driving the attack:
Improvised database dump with zero prior schema knowledge — The agent explored the database structure in real time rather than using a pre-written script.
A Chinese-language planning comment — Translating to "see what else we can do" — leaked into the command stream. This appears to be the agent's internal reasoning breaking through into execution.
Machine-readable command formatting — Every command used structured delimiters, bounded output caps, and discarded error streams. The agent was formatting its output for itself to parse, not for a human to read.
Adaptive command generation — Commands were composed in real time, adapting at each step to whatever the target revealed. The agent was feeding its own prior output into each next action, live and without human direction.
"We are watching attackers replace their scripts with AI." — Michael Clark, Sysdig TRT
Why This Changes Everything
This attack represents a paradigm shift in threat detection and response. Here's why:
Static Detection Is Now Obsolete
Traditional security rules built around specific command patterns or IP addresses are structurally inadequate against LLM-driven attacks. An AI agent rewrites its approach for every target, making signature-based detection largely ineffective.
Speed Amplification
The SSH bastion phase — credential replay, key retrieval, lateral movement, and database exfiltration — completed in under two minutes. A human operator performing the same steps would take orders of magnitude longer, giving defenders time to detect and respond.
Distributed Evasion
The agent's use of 12 different cloud API calls across 11 distinct IPs from Cloudflare Workers demonstrates a native understanding of distributed evasion that most human attackers spend years developing.
CVE-2026-39987: Already on CISA's List
The vulnerability used in this attack — CVE-2026-39987 — is already on CISA's Known Exploited Vulnerabilities catalog. Organizations running Marimo should patch to version 0.23.0 or later immediately.
Broader Implications for the Agent Ecosystem
This incident is not just a security story — it's an agent story. The same technology that powers beneficial AI agents — coding assistants, customer service bots, research tools — has now been weaponized by threat actors.
The Defensive Counterpart
On the same day Sysdig published its findings, GitHub's Agent of the Day blog showcased a Daily Security Red Team Agent — a Claude-powered workflow that runs nightly against GitHub Actions repositories, looking for backdoors, secret leaks, and supply-chain compromise. The agent opens up to five GitHub issues per run with the label security, red-team.
This dual-use reality — AI agents as both threat and defense — will define the next phase of the cybersecurity industry.
What Security Teams Should Do Now
- Patch Marimo immediately to version 0.23.0 or later
- Shift detection from static patterns to behavioral outcomes — monitor for credential access, database exfiltration, and unusual API call patterns rather than specific command signatures
- Audit AWS Secrets Manager access for any API keys that can be retrieved without MFA
- Monitor Cloudflare Worker egress for unusual data volumes or database connections
- Prepare for speed — assume that any breach can progress to exfiltration in under an hour
The Road Ahead
As Sysdig's report makes clear, this is not an isolated incident — it's a preview of the next generation of cyber threats. Attackers are not being replaced by AI; they are replacing their scripts with AI, gaining adaptability, speed, and evasion capabilities that were previously impossible to achieve at scale.
For the AI agent community, this is a sobering reminder that the same autonomy that makes agents powerful also makes them dangerous in the wrong hands. The security measures we build into our agent frameworks — sandboxing, rate limiting, human-in-the-loop checks, output verification — are not optional features. They are the last line of defense between beneficial AI and weaponized AI.
The age of autonomous AI cyberattacks is here. The only question is how quickly the defensive side can catch up.
Sources: Sysdig Blog - AI Agent at the Wheel, Breached.Company Analysis, GitHub Agent of the Day, AI Weekly
Cet article a été initialement publié sur The Agent Report.