AI Attacks Are Getting Faster and More Automated
Published: November 17, 2026
I run tigzig.com, an open platform with free data APIs and SQL access. Attacks on it have ramped up steadily since early this year, and security takes anywhere from a quarter to half of my time and AI token budget.
An AI team is now on round-the-clock surveillance with protocols for autonomous incident management, investigation, defensive action and escalation to me at different levels. I am a small tool builder, and I do this out of necessity. I can only guess how much harder it is for a large company.
Google's latest report projects attacks with little to no human involvement in 1 to 3 months, and sophisticated, fully autonomous ones in 3 to 6. Its 7-stage scale of AI attack automation puts attackers at stage 5 today.
The FBI says some intrusions with the ARTEX tool go from first target to privileged access in under three hours. ARTEX works with Claude, GPT and Chinese models.
Project Zero, the Google team that finds software flaws and reports them to the makers, will cut the time vendors get to fix a flaw from 90 days to 30 by 2028.
These attacks still use known weaknesses, such as compromised logins, unpatched systems and basic app flaws.
I come from a data science background and build AI tools. Building has become easy and validation is still hard, but security has been the hardest part for me.
A few quick points
▸ The FBI says ARTEX does best against systems that get less attention than the main app: partner and supplier portals, mobile backends, legacy services and third-party integrations.
▸ Google describes LLMjacking, where attackers break into cloud accounts and use the victim's AI model quota. In some cases single-day fraudulent charges passed $100K, and the worst reached $1M before it was stopped.
▸ After PHP released patches on 21 Oct, hundreds of claimed but non-functional exploit proofs, likely written with AI, were published, and defenders had to check them.
▸ Google expects spy groups, which must stay hidden, to use these noisy AI tools less. Cyber crime groups care less about hiding and should keep using them.
▸ I keep a 135-item security checklist for small tool builders at:
▸ Earlier related analysis: Sophisticated attacks no longer require sophisticated attackers, 21 Sep
tigzig.com/post/anthropic-threat-intelligence-cyber-sep2026
AI attacks are getting faster and more automated
Browse the slides or download the PDF
Full Deck Content (Text Format)
Text below was extracted from the source deck. Chart visuals stay in the PDF and as slide images above the post.
Slide 1
AI attacks are getting faster and more automated
Hackers have used AI for some time. What is new is how much of the attack AI now handles. ARTEX is the example.
What ARTEX is
An open-source security-testing tool that attackers have taken up. A person types a goal in plain words, and AI agents carry out the attack.
The FBI says
In some intrusions, ARTEX takes attackers from first target to privileged access in under three hours.
What it has done, and what it takes
Real damage. Data on about 67,000 people was taken from South Korean financial institutions. A China-aligned state group, per Microsoft, used ARTEX against more than 2,250 ConnectWise downstream tenants.
Little skill needed. People with minimal technical expertise can run complex attacks on many targets at once.
Many AI models plug in. Claude, GPT and Chinese models all connect. Chinese ones are most common among attackers, helped by low cost and few guardrails.
Write-ups on ARTEX since 30 October (selected)
30 Oct Google: ARTEX and Friends
6 Nov Microsoft: ARTEX abuse expands to the financial sector
11 Nov Palo Alto Networks Unit 42: ARTEX and AI-powered red teaming tools
13 Nov FBI: flash alert on ARTEX and Cisco Talos: ARTEX and beyond
SourcesFBI flash alert FLASH-20261113-001, 13 Nov 2026. Google Threat Intelligence Group, AI Threat Tracker, 16 Nov 2026.
Amar Harolikar · Decision Sciences & Applied AI17 November 2026 · 1 of 3tigzig.com
Slide 2
Google puts attackers at stage 5 of 7 on the way to full automation
Google's own scale of how attackers use AI, 16 November. Stages 1 to 5 have been seen. It projects stages 6 and 7 within 1 to 6 months.
1 Prompt-based use of LLMs End of 2025 to early 2026
Human-driven (chatbot)
2 Embedding LLM capabilities into malware Early 2026
Automated runtime call
3 Incorporating LLM-enabled automation into operations Early 2026 to Sep 2026
LLM-in-the-loop
4 Developing and experimenting with autonomous frameworks Mid 2026 to Sep 2026
Experimental autonomy
5 Multiple attack stages conducted by LLM-driven tooling in the wildNOW Current stage
Minimal human-in-the-loop
6 Multi-agent, unsophisticated attacks with little to no human interaction Near term: 1 to 3 months
No human-in-the-loop
7 Sophisticated, fully autonomous attacks Medium term: 3 to 6 months
Fully autonomous
Seen in the wild so far: ARTEX (FBI and Google), and CAIRN and PentAGI (Google). Solid boxes are observed, dashed boxes are Google's projection.
SourceGoogle Threat Intelligence Group, AI Threat Tracker, 16 Nov 2026, Figure 2. Redrawn here with Google's wording shortened in places.
Amar Harolikar · Decision Sciences & Applied AI17 November 2026 · 2 of 3tigzig.com
Slide 3
The time to fix a flaw is running out
Project Zero shortens its deadline Project Zero is Google's security research team. It finds software flaws, reports them to the maker and publishes them after a deadline. Days to that deadline: Before 13 Nov up to 120 days From 13 Nov 90 days From 2 Feb 2027 60 days From 3 Jan 2028 30 days
- Until now: 90 days to fix, plus 30 for users to update.
Patch to attack, almost at once. Google's bug team calls zero patch-to-exploit time "a reasonable, though imperfect, model" of attackers.
More flaws exploited. Disclosures and exploits nearly doubled from January to August 2026. Zero-days rose only marginally, from 8 a month in 2025 to 11 in 2026, before a spike to 22 in August.
No human in the loop. For the first time, Google found a tool that compares code before and after a patch, drafts an exploit and runs it.
AI helps defenders too. Project Zero now finds several flaws a week where it once found a few a year.
UK cyber agency (NCSC): AI-powered malware is "highly likely" in a ransomware attack on UK critical infrastructure by 2028.
What has not changed Google: the mechanics of intrusions "remain the same, such as using compromised credentials and targeting unpatched systems".
Not fully automated yet Google says these tools "currently trade stealth for velocity". WIRED: fully AI-powered ransomware attacks have "yet to emerge".
Earlier notes (tigzig.com ➜ Analysis)
21 Sep Sophisticated attacks no longer require sophisticated attackers
10 Sep Four warnings about AI and cyber risk arrived inside three weeks
This is my analysis of public sources, for information only. Care has been taken, but errors are possible. Use of AI. Parts of this work were prepared with AI under my direction. I reviewed, edited and approved every page. The analysis and conclusions are mine.
SourcesGoogle Project Zero, 13 Nov. Google Threat Intelligence Group, 16 Nov. UK NCSC, 3 Nov. WIRED, 14 Nov 2026.
Amar Harolikar · Decision Sciences & Applied AI17 November 2026 · 3 of 3tigzig.com
Working on something similar? How I work covers the rates, the availability and what I take on.