Automated AI cyberattacks in 2026: what Anthropic and Google actually reported
The headlines say AI is hacking on its own now. The original reports are more careful than that, and more useful. This post sticks to what Anthropic and Google's threat intelligence team (GTIG) published, keeps what they saw apart from what they only believe, and finishes with checks you can do this week. Where something is our own view, we say so.
Are AI agents really running cyberattacks?
Yes, in part. Attackers are using AI agents to speed up real intrusions, and a few documented cases are public. In the best-documented one, the AI did most of the routine work and humans chose the targets and approved the key steps. Google says it has not yet observed fully autonomous exploitation pipelines used against real targets (GTIG, September 2026).
Case 1: an espionage campaign run through Claude Code
Anthropic published this on 13 November 2025. It says it detected suspicious activity in mid-September 2025 and assessed with high confidence that a Chinese state-sponsored group was behind it. The group manipulated Claude Code and went after roughly thirty targets, including large technology companies, financial institutions, chemical manufacturers and government agencies. Anthropic says the attackers succeeded in a small number of cases (Anthropic).
- How much the AI did. Anthropic estimates it carried out 80 to 90 percent of the campaign, with humans stepping in at perhaps four to six decision points per campaign. At its peak the AI made thousands of requests, often several per second. That figure is Anthropic's own estimate.
- How the safeguards were got around. The operators broke the job into small tasks that looked harmless, and told the model it worked for a legitimate security firm doing defensive testing.
- Where it fell short. The model sometimes made up credentials, or claimed to have found secrets that were really public. Anthropic calls that an obstacle to fully autonomous attacks.
- What Anthropic did. It banned the accounts, notified the affected organisations, worked with authorities and improved its detection.
Anthropic calls this the first reported large-scale attack run with little human involvement. That is its own claim, and it adds that it can only see how its own tool was used.
Case 2: an exploit Google believes AI helped write
In its May 2026 report, GTIG said this was the first time it had identified a threat actor using a zero-day exploit that it believes was developed with AI. A group of cybercriminals planned a mass exploitation campaign around a Python script that bypassed two-factor authentication in a popular open-source administration tool. The flaw needed valid login details and came from a logic error. Google worked with the vendor to get it fixed and says its discovery may have stopped the mass use. GTIG says it has high confidence an AI model was involved, judging from the code itself, which had textbook-style documentation and an invented severity score. It did not think Google's own Gemini was used (GTIG, 12 May 2026).
The same report described PROMPTSPY, an Android backdoor first identified by security firm ESET. Google found that it sends a description of the screen to an AI model and carries out the taps and swipes the model sends back. Google disabled the related assets and found no copies on Google Play.
Case 3: credential theft run by AI agents
GTIG's September 2026 report describes a suspected financially motivated actor who broke into a victim's cloud infrastructure and set up a framework of AI agents. With an AI coding chatbot, a prompt and written agent instructions, the actor planned, built and ran a mass credential-theft campaign in under six hours and took thousands of third-party credentials. The agents handled the scanning, the troubleshooting and the IP rotation without manual work. In a related find, GTIG reached an exposed control server whose dashboard held more than 23,800 stolen secrets, cloud and AI API keys among them (GTIG, 9 September 2026).
The same report covers the software supply chain. A group GTIG tracks as UNC6780, also called TeamPCP, has been compromising packages on PyPI, npm and Docker Hub since March 2026. Its credential stealer, which GTIG names DUSTMAKER, takes tokens from GitHub Actions build runners so malicious packages look properly signed. It also hides files in folders used by AI coding tools, such as .claude, .vscode and .cursor. GTIG believes AI-assisted coding contributed to the large supply chain compromises of 2025 and early 2026. That is Google's assessment, not something it proved.
What is not proven
- GTIG says it has not yet observed fully autonomous exploitation pipelines used against targets in the wild, and describes a gradual maturing instead.
- Attribution to states and named groups is a judgement by the companies reporting.
- Figures like the 80 to 90 percent come from the organisation that published them, not an independent check.
What can a small business do about it?
The reports do not give small-business advice, so this is our reading of them plus the official guidance we have linked. What these cases have in common is ordinary weaknesses: stolen credentials, exposed keys and packages that were trusted too quickly. Eight things to check this week:
- Turn on multi-factor sign-in for email, billing and remote access first. CISA says any MFA beats none, and that phishing-resistant MFA is the standard to aim for (CISA).
- Look for keys sitting where they should not. Search your repositories and config files for tokens. If you find one, rotate it first. GitHub's guidance is that once a leaked secret is revoked, rewriting history is often unnecessary (GitHub docs).
- Give every key the least access it needs. Cloudflare's token guide shows scoping a token to specific accounts or zones, with an optional expiry (Cloudflare docs). Google's pipeline guidance also favours short-lived tokens over long-lived personal ones (GTIG, September 2026).
- Slow down new packages. Google's supply chain guidance suggests a release-age cooldown of at least 24 hours for npm and pnpm, and turning off install scripts with ignore-scripts=true (GTIG, July 2026).
- Pin exact versions. Google's pipeline guidance recommends exact versions with verified lockfiles instead of loose ranges like ^ and ~ (GTIG, September 2026).
- Look inside your AI tools' folders. After GTIG's finding, we would check .claude, .vscode and .cursor in any repository you clone or build from, for files you did not expect. That one is our suggestion.
- Check who controls what. The free Single Point of Failure Scanner shows which vendors run your email, DNS and sign-in.
- Review your app's login and privacy rules. If you run a Bubble® app, the 15 minute checklist is a good place to start.
Anthropic's own advice for security teams is to use AI on the defending side too, for threat detection and incident response. For most small businesses that starts with switching on the alerts your providers already offer, and actually reading them.
Questions people ask
Can AI hack a business with no human involved?
Not in the published cases. Humans still picked targets and approved key steps. Anthropic estimates AI did 80 to 90 percent of the work in one campaign, and Google says fully autonomous exploitation has not been seen in the wild.
Is AI inventing new kinds of attack?
The cases above are stolen credentials, a software flaw and poisoned packages. Those are not new. In our reading, what AI changed in these reports is the speed and the amount of work one attacker can get through.
Should a small business be worried?
We would call it a reason to do the basics properly rather than a reason to panic. All eight checks above use free or built-in settings.
Sources
- Disrupting the first reported AI-orchestrated cyber espionage campaign, Anthropic, 13 Nov 2025
- GTIG AI Threat Tracker: Adversaries Leverage AI for Vulnerability Exploitation, Augmented Operations, and Initial Access, Google Threat Intelligence Group, 12 May 2026
- GTIG AI Threat Tracker: From Prompting to Autonomy, the Evolution of Adversarial AI, Google Threat Intelligence Group, 9 Sept 2026
- Batten Down Your Packages: Mitigation Guidance for Supply Chain Compromise, Google Threat Intelligence Group and Mandiant, 31 July 2026
- Proactive Defense: Hardening Code Pipelines and CI/CD Infrastructure, Mandiant, 25 Sept 2026
- More than a Password (multifactor authentication), CISA
- Removing sensitive data from a repository, GitHub Docs
- Create API token, Cloudflare