Deep diveCyber Defense

AI cyber threats: what Anthropic, Google, OpenAI and the UK NCSC have reported

Official reporting from Anthropic, OpenAI, Google's Threat Intelligence Group and the UK National Cyber Security Centre points the same way: attackers use AI to work faster and at larger scale, mostly through familiar methods. Human operators still direct campaigns. The clearest shift is economic, and the sharpest operational effect is shorter time between disclosure and exploitation. This page summarizes what each source says and where its limits lie.

Reviewed 6 min read

On this page
  1. What the reports agree on
  2. What each source documents, by date
  3. Terms that recur in provider threat reports
  4. The economics behind the reports
  5. Limits of what these reports can show
  6. Where the evidence points for defenders
  7. What to watch in the next reports
  8. Questions and answers
  9. Sources

What the reports agree on

Four organizations publish regular, primary reporting on this topic: AI developers that detect misuse on their own platforms, and the UK government's national cyber authority. Their vantage points differ, which makes agreement among them more informative than any single report.

Anthropic's report of September 2026, covering December 2025 to August 2026, says none of the operations it described relied on a wholly novel technique and that the change lies in the economics1. It says AI has inverted the cost back onto defenders, and that the main distinguishing feature between classes of actors is no longer sophistication but intent1.

The NCSC's 2026 annual review reaches a similar view2. It judges that AI speeds up and scales existing methods, that human operators remain central, and that defenders face a race to shorten patch deployment timelines2. Google's analysts made the same point earlier: in January 2025 they reported productivity gains but not novel capabilities3.

What each source documents, by date

SourceDate and what it reportsLimit to keep in mind
NCSC assessmentPublished 24 January 2024: AI will almost certainly increase the volume and heighten the impact of cyber attacks, with phishing and social engineering gaining most4.A forward-looking judgment, not a count of incidents.
Google GTIGPublished 30 January 2025: threat actors gained productivity from AI but were not yet developing novel capabilities3.Observations of misuse of one provider's models.
NCSC forward assessmentPublished 7 May 2025: a digital divide is expected between organizations that keep up and those that do not; AI-assisted vulnerability research and exploit development is the most significant development; fully automated advanced attacks are unlikely by 20275.Judgments with stated confidence.
AnthropicPublished 27 August 2025: an extortion campaign against at least 17 organizations, with ransoms that sometimes exceeded $500,000, and a ransomware-as-a-service actor selling packages for $400 to $1,2006.Cases Anthropic detected and disrupted on its platform.
OpenAIPublished 7 October 2025: reports having disrupted over 40 networks and says threat actors bolt AI onto old playbooks to move faster7.Cases seen on OpenAI's services only.
Google GTIGPublished 6 November 2025: first use of just-in-time AI inside malware, including families named PROMPTFLUX and PROMPTSTEAL; actors posing as capture-the-flag participants or researchers to get past guardrails; a maturing underground marketplace8.Early and experimental malware use.
Google GTIGPublished 13 February 2026: a model distillation campaign of over 100,000 prompts, plus malware families HONESTCUE and COINBAIT; GTIG says it has not yet observed breakthrough capabilities9.Reflects what GTIG could observe at that date.
AnthropicPublished September 2026: an AI-run supply chain campaign in which agents performed nearly all of the work and collected over 2,100 Azure AD token sets from more than 40 tenants in about 34 hours1.Humans still made key decisions such as target selection1.

OpenAI's page could not be opened directly in the review behind this page; its entry rests on a search extract of the publisher's own text and should be verified before quoting.

Terms that recur in provider threat reports

Distillation campaign
Sending a model very large numbers of prompts to copy its behavior; Google's analysts described one of over 100,000 prompts9.
Just-in-time AI in malware
Malware that calls a language model while it runs, rather than carrying fixed logic; Google's analysts reported the first such use in November 20258.
Ransomware-as-a-service
A model in which an operator sells ransomware packages to others; Anthropic reported packages priced from $400 to $1,2006.
Digital divide
The NCSC's term for the widening gap between organizations that keep pace with AI-enabled threats and those that fall behind5.
Shadow AI
AI tools used by staff without approval, which the NCSC says reduces an organization's visibility of its data2.

The economics behind the reports

Anthropic's most useful framing is economic. When AI agents perform most of an intrusion's repetitive work, as Anthropic reports for a supply chain campaign that gathered more than 2,100 token sets from over 40 tenants in about 34 hours1, the cost of each additional target falls. Defenders then face more attempts, not necessarily cleverer ones.

The NCSC places the same pressure on patching. Its 2025 assessment says the gap between disclosure and exploitation has shrunk to days, and that AI-assisted vulnerability research and exploit development is the most significant development it tracks5. The 2026 annual review adds that unapproved AI tools in organizations reduce visibility, which complicates defense further2.

Attackers also target AI systems themselves. Anthropic reports an actor it tracks as GTG-50020 injecting instructions into an AI vendor's evaluation sandbox to obtain production API keys1. The prompt injection defense guide covers the controls for that class of attack.

Limits of what these reports can show

Where the evidence points for defenders

  • If

    Your exposed systems depend on patching cycles measured in weeks.

    Then

    Set a shorter target for internet-facing and actively exploited vulnerabilities, and measure it.

    The NCSC reports that time between disclosure and exploitation has shrunk to days5.

  • If

    Staff or developers use AI tools or keys you have not inventoried.

    Then

    Inventory them and treat AI keys and agent integrations like production credentials.

    Anthropic advises that treatment, and the NCSC says unapproved AI use reduces visibility12.

  • If

    Your help desk or finance team can be persuaded by a convincing message.

    Then

    Add out-of-band verification for credential resets and payment changes.

    The NCSC expects phishing and social engineering to gain most from AI4.

  • If

    You buy AI access through unofficial resellers or shared accounts.

    Then

    Move to authorized channels with named owners and revocable keys.

    Anthropic's September 2026 report advises buying AI access only through authorized channels1.

What to watch in the next reports

Three questions are worth tracking as new reports appear. First, whether providers begin to document capabilities that were absent before, which is the threshold Google's analysts say they have not yet seen crossed9. Second, whether just-in-time AI inside malware, first documented in November 20258, moves from experiments to routine use. Third, whether the digital divide the NCSC describes shows up in measurable differences between organizations that patch quickly and those that do not5.

For decisions that follow from this evidence, see the AI security strategy page and the guide to AI red teaming.

Questions and answers

Are attackers using AI to invent new kinds of attack?

Official reporting says mostly not yet. Anthropic says none of the operations in its September 2026 report relied on a wholly novel technique1, and Google's analysts said in February 2026 they had not yet observed breakthrough capabilities9. What has changed is speed, scale and cost.

What is the biggest operational effect of AI on attackers?

The NCSC identifies AI-assisted vulnerability research and exploit development as the most significant development, and says the time between disclosure and exploitation has shrunk to days5. For defenders, that makes patch speed and exposure management the most direct response.

Can AI agents run an intrusion on their own?

Anthropic reports a campaign in which agents performed nearly all of the work, though humans still made key decisions such as target selection1. The NCSC judged in May 2025 that fully automated end-to-end advanced attacks are unlikely by 20275. Plan for heavy automation with human direction.

Which sources should a board rely on for AI threat claims?

Prefer primary reporting from government authorities such as the NCSC and from providers' own threat intelligence teams, and read each one's stated limits. Treat consultancy surveys and vendor marketing claims as secondary. When a figure matters to a decision, trace it to the original report.

Sources

  1. Detecting and countering misuse of AI: September 2026 — Anthropic · checked 19 November 2026
  2. Annual Review 2026: the impact of AI on the cyber threat — UK National Cyber Security Centre · checked 19 November 2026
  3. Adversarial Misuse of Generative AI — Google Threat Intelligence Group · checked 19 November 2026
  4. The near-term impact of AI on the cyber threat — UK National Cyber Security Centre · checked 19 November 2026
  5. Impact of AI on cyber threat from now to 2027 — UK National Cyber Security Centre · checked 19 November 2026
  6. Detecting and countering misuse of AI: August 2025 — Anthropic · checked 19 November 2026
  7. Disrupting malicious uses of AI: October 2025 — OpenAI · checked 19 November 2026
  8. GTIG AI Threat Tracker: Advances in Threat Actor Usage of AI Tools — Google Threat Intelligence Group · checked 19 November 2026
  9. GTIG AI Threat Tracker: Distillation, Experimentation, and Integration of AI for Adversarial Use — Google Threat Intelligence Group · checked 19 November 2026

More in Cyber Defense

Back to Cyber Defense

Next step

Turn threat reporting into patch and detection targets

Send us your current remediation times for exposed systems and how you detect unusual credential use. We will outline which targets to tighten first and how to test that detection meets them.

Review our targets

Your starting point

What would you like to move forward?

Start with the outcome that feels closest. We'll help give the next step a useful shape.

Let's find your next move

One conversation. A clearer direction.

Tell us what you want to improve. A few sentences are enough to start.

shayan@coldai.org