The AI agent now attacks independently
„Attackers use AI agents to fully automate cyberattacks from reconnaissance to data theft, enabling smaller actors to run state-level operations.”
Counter-narrative
The label 'first fully autonomous attack' keeps coming up, but according to several reports, AI agents are not yet capable of fully autonomous attacks: they are accelerator tools, and human involvement is also required. This is represented by some reports and critics of the Sysdig case.
Voices
Anthropic, Google, Microsoft, FBI, Sysdig, professional security press
Analysis
According to the narrative, agents have appeared in every stage of the attack chain, and the barrier to entry has collapsed. It is fed by AI labs' threat reports and security firms' case write-ups, which award the label "first" several times in a row. Labs have an interest in acting as responsible gatekeepers, and security vendors in selling AI-based defense. The weak point is that there is no commonly accepted definition of the boundary between "autonomous" and "AI-assisted." The debate over the Sysdig case and the contradictory February reports show how fluid the basic concept is.
What's at stake: The stake is whether defense aligns with real automation or a marketing concept.
Narrative sub-threads
State-sponsored espionage campaigns with autonomous agents
Actors linked to China, Russia, and other countries are running espionage operations against government and corporate targets using AI agents, from reconnaissance to data collection. According to Anthropic, China-linked attackers used agents against approximately thirty large corporations and government agencies. In Taiwan, autonomous agents breached 85 government accounts. CrowdStrike states that Chinese state hackers targeted more than 200 organizations with ARTEX.
It concerns state-sponsored or state-linked espionage groups and government victims, not financially motivated crime. The evidence consists of vendor threat reports and government statements.
▲ at peak · evidence: strongSmall players at state scale
Individual hackers and small criminal groups are breaching hundreds of organizations using commercial AI tools, sometimes within hours. According to Anthropic, this previously required state-level capabilities. Examples: 600 firewalls, the PaperCut servers of 395 organizations, 2,250 Microsoft 365 tenants breached in hours, and 150 GB of Mexican government data.
The actor here is an individual or small group. The mechanism involves the abusive use of generative AI such as Claude, Cursor, DeepSeek, or others. The evidence points to specific mass intrusions.
→ lasting · evidence: moderateAI-written malware and automated ransomware
Attackers use AI to generate malware and exploits, such as VoidLink, the Hive Slopoly malware, and a zero-day exploit detected by Google. According to Sysdig, an agent fully automated a ransomware attack; this claim is disputed. One report states that agents can reduce ransomware intrusion time to under ten hours.
The focus is on automating attacker tools and ransomware chains, not the identity of the actor. Evidence includes malware analysis and incident descriptions.
↗ rising · evidence: moderateAuthority alerts: ARTEX and forecasts
The FBI and Google warn that attackers using the open-source ARTEX testing tool and other AI platforms can gain privileged access within three hours in some breaches. Google says fully autonomous attacks could appear within three months. The UK NCSC considers AI-based ransomware attacks on critical infrastructure extremely likely by 2028.
It relies on law enforcement, regulatory, and manufacturer warnings and forecasts, not on a single incident. The ARTEX tool and agents building exploits from patches are at the center.
↘ fading · evidence: mediumContested autonomy and excessive fears
Sources say agents are not yet capable of fully autonomous attacks. The 'first' AI-driven cases are disputed regarding both human involvement and attribution. According to VulnCheck, only 1% of vulnerabilities discovered by AI were exploited in live attacks.
It carries counter-evidence and a skeptical assessment, questioning the evidence of other narrative branches.
Sub-thread stories
State-sponsored espionage campaigns with autonomous agents
Small players at state scale
AI-written malware and automated ransomware
Authority alerts: ARTEX and forecasts
Contested autonomy and excessive fears
Who is affected and how
- security leader
- Expect intrusion times in the order of hours: according to an FBI warning, attackers using ARTEX AI agents can obtain elevated privileges within three hours. Detection and isolation SLAs should be aligned with this, not with the 'first' label of campaign news.
- security operations
- Agent attacks occur at machine speed: Sysdig reports a 31-second attack, while another report details 2,250 Microsoft 365 tenants compromised within hours. Alerts should be built for rapid, mass authentication and reconnaissance sequences, and the SOC should be tested on its response time.
- public-sector decision-maker
- Based on government cases in Mexico, Taiwan, and Thailand, administrative systems and accounts are high-priority targets. Enforcing MFA and access control is more important than AI-specific regulation.
- journalist and researcher
- For every claim of a 'first autonomous attack', ask who determined autonomy based on what evidence and whether there is independent confirmation.
Hungary
No or minimal domestic trace during the period under review: the list contains no [HU] story. It is worth monitoring whether domestic incidents are retroactively labeled with AI tags without evidence.
Weekly trend
Turning points
Related narratives in the News magazine
Supporting stories
Machine analysis based on story titles and summaries, with a verification pass. The source of the facts is the article linked at each story.
What do you make of this piece?
The split shows after 20 votes.
Why? Your answer is not shown publicly; it helps the editors.