How to Detect and Defend Against Autonomous AI Cyberattacks

Summary
The article explains how AI agents could automate intrusion tasks, citing Anthropic's estimate that an agent handled 80–90% of tactical work in a state-sponsored campaign, and recommends behavior-based detection, identity monitoring and machine-speed defense.
Key points
- The article says Anthropic disclosed a late-2025 state-sponsored espionage campaign in which an AI agent performed an estimated 80–90% of tactical work across roughly 30 targets, with humans intervening at a few decision points.
- It describes agents using an observe-plan-act-evaluate loop and sub-agents connected to tools; defenders should look for sequences and tempo, not only individual events.
- Suggested detections include rapid sequential discovery commands, bursts of authentication attempts across accounts, and processes spawning many reconnaissance tools in seconds.
- Recommended monitoring includes anomalous API use, unusually fast sessions, unexpected server calls to LLM endpoints, and honeytokens or decoy credentials.
- The article says CERT-UA's LameHug, linked to APT28, used a hosted LLM to generate reconnaissance commands; the PromptLock proof of concept demonstrated ransomware generating Lua scripts with a local model.
- Preparation advice includes mapping detection coverage, patching exposed systems, enforcing phishing-resistant MFA, rotating secrets, removing standing privileges, and inventorying internal AI agents and MCP servers.
- It recommends automating detection engineering and alert enrichment while keeping human approval for destructive response actions.
Article Details
- Defense Focus
- Detect autonomous attacks through behavioral patterns, event sequence, and activity velocity rather than relying primarily on campaign-specific indicators.
- Detection Methods
- Detect rapid sequential discovery commands and bursts of authentication attempts across many accounts.
- Correlate low-severity actions by identity and time; the article gives ten actions within 60 seconds as an example signal.
- Monitor for new processes spawning numerous reconnaissance tools in seconds.
- Hunt for anomalous API usage, sessions operating at impossible speed, and unexpected calls to LLM endpoints from servers.
- Look for outbound traffic to LLM APIs from non-user hosts, embedded API keys in binaries, local inference runtimes on endpoints, and scripts that appear freshly generated.
- Use honeytokens and decoy credentials to detect agents validating discovered credentials.
- Data Sources
- Identity and service-account activity
- Authentication events
- API usage and LLM endpoint traffic
- Process creation and endpoint telemetry
- Outbound network traffic
- Binary contents for embedded API keys
- Rule Types
- Sigma detection rules
- Behavior-based detections
- Time-windowed sequence and velocity correlation
- Platforms
- SIEM
- EDR
- Data lakes
- Defensive Actions
- Map detection coverage to relevant ATT&CK behaviors using available log sources and address coverage gaps.
- Continuously assess external attack surface and rapidly patch internet-facing systems.
- Enforce phishing-resistant MFA, rotate secrets, and remove standing privileges.
- Inventory internal agents, MCP servers, and LLM integrations; assess them for prompt-injection and hijacking risk.
- Run AI-assisted adversary-emulation purple-team exercises and measure time to detect.
- Maintain detection-as-code pipelines with automated testing and deployment; refresh detection content as behaviors change.
- Use AI agents to enrich and correlate alerts, but require human approval for destructive response actions.
MITRE ATT&CK
T1003 · OS Credential DumpingCredential dumpers are listed as tooling wired to attack sub-agents, which also harvest and validate secrets.T1110 · Brute ForceThe article identifies burst authentication attempts across many accounts as behavior to detect in agent-driven attacks.T1595 · Active ScanningThe described autonomous agent enumerates exposed services during reconnaissance using scanners.
Threat Actors
Malware
LAMEHUGCERT-UA’s LameHug, linked to APT28, used a hosted LLM to generate recon commands on the fly, and the PromptLock proof of concept showed ransomware writing its own Lua scripts via a locally run model.PromptLockCERT-UA’s LameHug, linked to APT28, used a hosted LLM to generate recon commands on the fly, and the PromptLock proof of concept showed ransomware writing its own Lua scripts via a locally run model.
Vendors
Products
SOC Prime PlatformThreat Detection Marketplace on the SOC Prime Platform delivers curated, ATT&CK-mapped rules as emerging threats surface, so coverage evolves without waiting on a backlog.Threat Detection MarketplaceThreat Detection Marketplace on the SOC Prime Platform delivers curated, ATT&CK-mapped rules as emerging threats surface, so coverage evolves without waiting on a backlog.Uncoder AISOC Prime’s Uncoder AI does exactly this, turning CTI into ready-to-deploy detections and translating rules across dozens of SIEM, EDR, and data lake languages.
