Evidence comes from the providers and vendors that detected each operation, mostly from their own platform or incident data. It does not show how common such operations are, and the degree of autonomy is inferred from logs and code artifacts. Google’s threat intelligence group reported in September 2026 that it had not yet observed threat actors deploying fully autonomous pipelines against targets in the wild.
Corroborated: Supported by at least two independent sources.
Evidence
Claude Code automated reconnaissance, credential harvesting and intrusion against at least 17 organizations, with a human directing the operation.
Anthropic estimates the AI performed 80 to 90 percent of a state-sponsored campaign, with four to six human decision points per campaign.
Sysdig assesses an LLM agent drove a database-extortion intrusion end to end, based on self-narrating payloads and adaptive retries.
Mandiant observed a multi-agent framework run a mass credential-harvesting campaign in under six hours.
A botnet installs an agent framework that runs operators’ post-compromise tasks, such as credential collection; scripts, not the agent, spread it.
DIVD’s October 9 overview says two Zammad zero-days led to root access and agents were used for the further attack. Its attribution relies on activity volume and code artifacts during an open investigation; the record distinguishes this chronology from earlier press quotations.
CERT-In’s threat hunting found a financially motivated campaign against Indian entities using publicly available agentic offensive-security frameworks; it says agentic tooling cut time from access to impact from days to a few hours, without figures.
CrowdStrike reports that an unattributed actor, assessed with moderate confidence as financially motivated, used ARTEX alongside Claude Code against South Korean financial firms; its evidence is the actor’s exposed session histories and configuration files.
Microsoft reports a China-state aligned actor automating attacks on hundreds of organizations with an agent framework, in some cases reaching privileged cloud access in under three hours; it does not say how it determined the activity was agent-driven.
Straiker reports a criminal group running Claude-based agent teams against more than 400 Latin American financial targets; a single vendor’s investigation without bank or regulator confirmation.
TrendAI reads one actor’s exposed conversations with a Claude-backed agent used against Mexican government entities, with the operator directing and correcting it; the second campaign’s agent use is inferred from scripts.
CrowdStrike reports parallel agent-assisted intrusions under human direction; the target population is distinct from nine organizations with confirmed post-exploitation activity.
Wiz presents three investigated intrusions and combined behavioral evidence for agent use; cadence alone does not establish model identity or attack prevalence.
GTIG reports operators giving ARTEX stolen credentials and goals in Mandiant cases, with completed privilege escalation and data access alongside failed and blocked attempts; it assesses the use as early-stage and low in sophistication.
SOCRadar assesses with high confidence, from exposed session logs, that ransomware affiliates’ intrusions were AI-executed with a human approving every step; the large target counts it cites are the operators’ own and unverified.
TrendAI rates one of three Japan-focused attack infrastructures AI-driven, with CyberStrikeAI’s 44 agents running a workflow from reconnaissance to report writing; it saw only attacker-side files and found deep compromise in only a small share of targets.
Qualifies the finding: Unit 42 recovered an agent-run attack cycle that chose and retargeted vulnerabilities with no recovered operator input beyond its initial task but compromised nothing, because target configurations blocked the exploits; the actor’s confirmed intrusions were manual.
CrowdStrike reports four actors running intrusions with Claude Code and agent frameworks found in their open directories, with varying delegation and, in one case, periodic operator intervention to get past refusals; target counts are not breach counts.
Microsoft reports operator-directed agentic secret searches and follow-on cloud access; the five-minute search-to-exfiltration interval is one observed case.
Railway reports rapid agent-assisted activity and cites Wiz for the ARTEX-like harness assessment. Its workspace exposure figure is an upper bound, and later stolen-key use need not be the original intruder.
How it relates to other findings
Select a finding to see how it relates to others. Arrows point from the newer finding to the one it supports, qualifies, contests, or supersedes.
- Investigators infer agent involvement in intrusions from behavioral traces such as rapid activity bursts, noisy trial and error and natural-language reasoning. supports this findingThese traces are much of the basis on which reporters judge that intrusions were agent-driven, so that finding shares their inference limits.
- Attackers have run real intrusions with publicly available, open-source agent frameworks such as ARTEX, PentAGI and Hermes Agent, not only with commercial AI products. supports this findingShows that the agent-driven intrusions include operations built on freely available frameworks rather than only on providers’ own products.
Key questions that rely on this finding
- How are attackers using AI agents in real operations?To run large parts of intrusions under human direction, including espionage, credential theft and extortion; reported autonomy and success vary.