Reports of autonomous AI agents conducting live intrusions indicate a shift beyond research demonstrations into operational use. Our review also examines AI-assisted vulnerability discovery and attacks targeting the infrastructure supporting AI systems.
The Signals
-
01
Four threat actors ran autonomous AI agents as live intrusion tooling, and Australia's national cyber agency named AI as a key accelerant in the same week.
-
02
The behaviour is cheap and bounded mainly by the price of compute; one reported campaign cost about US$50 per break-in.
-
03
The vulnerabilities themselves are increasingly found with AI, PHP2Shell among them; those PHP flaws are now listed by CISA as actively exploited, though the attackers show no AI of their own.
-
04
The AI stack itself is now a target: a poisoned build of the LiteLLM AI gateway set out to steal cloud and developer credentials before it was caught.
01 / Threat activity
AI agents are crossing from research demos into intrusion tooling
Autonomous AI agents are moving from proof-of-concept to operational attack tooling. Last week, the Australian Signals Directorate named the shift in its annual threat assessment.
CrowdStrike reported four distinct threat actors using agentic AI to run intrusions (SecurityBrief). The activity ran between May and August 2026 with minimal human direction. In one case an agent worked for roughly 12.5 minutes after a single human prompt; in another, six agents ran in parallel for about four hours. One actor compromised at least 15 Argentine healthcare organisations, and a separate intrusion reached 726 tenant databases. CrowdStrike’s OverWatch team says it recorded three times as many agentic intrusions so far in the second half of 2026 as in the first.
The actors slipped past the models’ safety controls with simple tactics. According to CrowdStrike, they presented the work as an authorised penetration test, planted files that marked targets as approved, and rephrased or restarted a session after a refusal; one operator ran a modified build of a coding agent. The attack tooling was itself exposed: researchers at NorForge, reported by Cyber Press, found at least 17 publicly accessible control panels for the open-source agent tool Cairn that had been fed malicious injected tasks, some delivering a remote-access implant and adding attacker SSH keys, turning autonomous intrusion agents against their operators.
For now the limit on these attacks is their running cost. CrowdStrike calls token budgets a key constraint on how far an agent can go, which ties the volume of autonomous intrusion to the price of compute. That price is unsettled: in the same week Anthropic filed to list on Nasdaq, and banks and Blackstone lined up about US$63 billion to finance Broadcom AI chips for Anthropic. The Cairn campaign put today’s figure in plain terms, at a reported US$50 in tokens per successful break-in.
Defenders are running the same kind of agent. Causal Security’s “Cipher” found and validated four previously unknown flaws in the open-source EPICS industrial-control toolkit in under four hours, one rated critical at CVSS 9.8, and a CISA advisory and fix followed.
The national picture matched the incidents. The ASD’s Annual Cyber Threat Report 2025–26, released this week, logged more than 4,400 notifications of potentially hostile activity, attributed state-sponsored targeting of Australia and its allies to China, Russia and North Korea, and named AI as a factor accelerating vulnerability discovery and exploitation.
These behaviours can be measured. An agent chaining low-privilege access into administrative control, or talking its way past a safety control by posing as an authorised tester, can be observed and graded. Running a cyber-AI agent against a real, contained target and grading every action it takes is what FIFTH DOMAIN’s XORCISE evaluation engine does. We outline the wider assurance case in our Insights Report on the emerging assurance challenge.
02 / Vulnerability discovery
AI is finding vulnerabilities in bulk
Disclosed software vulnerabilities are piling up, and AI is part of why. Google’s Threat Intelligence Group counted monthly CVE disclosures roughly doubling in 2026, from about 5,045 in January to 10,740 in August. Flaws it judged likely to have been found with AI reached remote code execution about half the time, against 26 per cent for the rest, though only a small share of all disclosures, roughly one in 431, were seen exploited in the wild.
PHP2Shell was the clearest case. Two unauthenticated remote-code-execution flaws in PHP core, CVE-2026-109618 and CVE-2026-109619, both rated CVSS 9.2, were found in August by separate teams at Calif and Wiz, each using AI tools including frontier models; OpenAI said every group that reported the bug around that time had used AI. TechCrunch noted the same flaws had been judged not practically exploitable about two years earlier.
Exploitation followed within days of disclosure. PHP shipped the fix on 21 October, a day early after the details leaked; Amazon had logged an exploitation attempt four days before disclosure, and after release Perplexity researchers recorded 35 successful shell executions against honeypots. CISA listed both flaws as actively exploited on 26 October, and one monitoring group counted more than 50,000 exploitation attempts from over 600 addresses in 70 countries. PHP runs roughly 70 per cent of websites (W3Techs), and more than 5.1 million PHP hosts are exposed online. The attackers have not been identified, and nothing so far shows they used AI; AI’s role in PHP2Shell has been in discovery, with no sign it drove the exploitation.
The same pattern is turning up across the stack. Researchers at Xbow used OpenAI’s GPT-5.6 Sol to chain three Chrome zero-days into a working exploit affecting around a billion users on Xbow’s estimate, earning Google’s top reward for a Chrome full-chain exploit, US$250,000, for about US$40,000 in model costs. A BeyondTrust flaw found by the Hacktron AI agent was exploited within four days of disclosure. Google’s analysts warn that attackers could turn the same models on patched bugs to weaponise them faster, though they present that as a risk rather than something they have measured.
03 / Supply chain & infrastructure
The AI stack becomes a target
Attackers are now going after the tools used to build AI. On 21 October a malicious version of LiteLLM, a widely used open-source AI gateway, was published to PyPI and public container registries, Microsoft reported. The package ran its payload at Python startup, even for projects that never imported it, and set out to steal credentials for cloud accounts, AI assistants like Copilot, MCP configurations, Kubernetes, Docker and SSH, encrypting the haul and sending it to a lookalike domain. The maintainers said the release workflow had been hijacked with stolen GitHub credentials; Microsoft did not determine the initial access or name who was responsible.
Vulnerabilities in AI software are mounting more broadly. Google’s Threat Intelligence Group has tracked 2,076 of them since the start of 2025, with about half of this year’s in agent-building frameworks such as Flowise and Langflow and several hundred more in inference servers like vLLM and Ollama. The malicious LiteLLM build was live for only a few hours, and its reported download figure of about 49,000 is unconfirmed.
Mission assessment
What the evidence indicates
Reporting this week points to a broadening role for AI across cyber operations, including autonomous intrusion activity, vulnerability discovery and defensive security research. The incidents demonstrate varying degrees of AI involvement, from direct execution of attack sequences to AI-assisted identification of vulnerabilities subsequently exploited through conventional means.
The activity also highlights risks within AI infrastructure itself, particularly software supply chains and agent management interfaces. While these developments indicate an expanding operational role for AI, the evidence does not yet establish the scale, consistency or effectiveness of autonomous cyber operations across the wider threat landscape.
Continued observation will be necessary to assess how these capabilities evolve, their operational limitations and their implications for cyber defence and assurance.
Sources
- CrowdStrike finds AI agents running autonomous cyberattacks. SecurityBrief Australia, 28 Oct 2026. securitybrief.com.au
- Hackers hijack AI agents (Cairn / NorForge). Cyber Press, 27 Oct 2026. cyberpress.org
- AI agent finds four severe industrial-control bugs in under four hours (EPICS, ICSA-26-300-01). The Register, 27 Oct 2026. theregister.com
- Australia warns of AI-powered cyberattacks (ASD Annual Cyber Threat Report 2025–26). Euronews, 27 Oct 2026. euronews.com
- Anthropic files S-1 for Nasdaq IPO. Yahoo Finance, 26 Oct 2026. finance.yahoo.com
- Banks and Blackstone line up $63B to finance Broadcom chips for Anthropic. Reuters via The Star, 27 Oct 2026. thestar.com.my
- PHP2Shell: AI-assisted discovery of PHP core RCE flaws. Wiz, 21 Oct 2026. wiz.io
- AI tools discovered the PHP vulnerability that could take down 70% of the internet. Fortune, 21 Oct 2026. fortune.com
- CISA confirms PHP2Shell vulnerabilities now exploited in attacks. BleepingComputer, 27 Oct 2026. bleepingcomputer.com
- US government says hackers are now exploiting critical PHP web bugs. TechCrunch, 27 Oct 2026. techcrunch.com
- Researchers used GPT-5.6 Sol to build a billion-user Chrome exploit chain. The Register, 14 Oct 2026. theregister.com
- Vulnerability discovery and exploitation trends in the AI era (Google GTIG). Help Net Security, 1 Oct 2026. helpnetsecurity.com
- Inside the LiteLLM 1.107.0 supply-chain compromise. Microsoft Security Blog, 23 Oct 2026. microsoft.com