Threat briefing · Updated October 2026
In the past 18 months AI has found thousands of unknown vulnerabilities, written exploits used in the wild, run espionage campaigns and broken into real companies that never noticed. At the same time, AI is writing half of new code, and regulators have started to enforce.
What changed
Each one maps to a part of what we do. Most companies are exposed on more than one.
Models find vulnerabilities, write exploits and run campaigns at machine speed. Breakout can take seconds, not days.
How we helpInsecure defaults, leaked keys and invented packages ship straight to production, often from teams with no security review.
How we helpCopilots and agents read untrusted content and hold real permissions. One poisoned document can turn them against you.
How we helpThe EU Cyber Resilience Act and AI Act, South Korea’s AI Basic Act and new data laws turn security into a legal duty.
How we helpIncident file
Public, sourced events, newest first. Each one shows what it means for your company and which of our services addresses it.
Sep 2026
Under the Cyber Resilience Act, manufacturers selling products with digital elements in the EU must now report actively exploited vulnerabilities and severe incidents through ENISA’s Single Reporting Platform. Full obligations follow in December 2027.
What it means for you: If you sell software into the EU, you need a working vulnerability-handling process now, not in 2027.
Aug 2026
A self-spreading worm published 1,684 poisoned versions across 420 npm packages. It harvested GitHub, cloud, Kubernetes and private-key credentials, and left Claude Code and VS Code configuration files in repositories as another way to run its payload.
What it means for you: Your dependencies and your AI tool configuration are now part of the attack surface.
Jul 2026
OpenAI disclosed that models escaped a test environment and breached Hugging Face using a zero-day and exposed credentials. Days later, Anthropic reported its models had reached production systems at three companies during evaluations. None of the three had detected it. Analysis of the incidents found no novel exploits: the routes in were weak passwords, unauthenticated endpoints and over-privileged accounts.
What it means for you: Low-priority findings left open for years are exactly what autonomous attackers chain together.
Jul 2026
Veracode found AI now authors roughly half of committed code in organisations that use it, yet AI-generated code passed security tests in only 56% of tasks. For cross-site scripting the pass rate was 15%; for log injection, 12%.
What it means for you: Every AI-assisted team needs automated security checks between the assistant and production.
May 2026
Researchers found 380,000 publicly accessible assets on AI app-building platforms. About 5,000 held sensitive corporate data, from financial records and clinical trial details to patient conversations, often created by staff without IT approval.
What it means for you: Shadow AI is now shadow software. You need to know what your people have built and published.
May 2026
Google’s threat intelligence team found a criminal group using an AI-generated exploit for a two-factor authentication bypass in a popular open-source admin tool, prepared for mass exploitation. Tell-tale signs included a hallucinated CVSS score in the code.
What it means for you: The time between a flaw existing and being exploited is collapsing. Patch prioritisation has to keep up.
Feb 2026
CrowdStrike reported adversaries injecting malicious prompts into legitimate AI tools at more than 90 organisations to generate commands for stealing credentials and cryptocurrency, alongside an 89% rise in AI-enabled attacks.
What it means for you: Every AI tool connected to your data or systems needs testing and monitoring like any other privileged system.
Nov 2025
Anthropic disrupted a state-sponsored campaign that used an AI coding agent against roughly 30 organisations, including tech companies, banks, manufacturers and government agencies. The AI carried out 80 to 90% of the work, at thousands of requests, often several per second.
What it means for you: Attacks that once needed a skilled team can now run at machine speed with a handful of people.
Aug 2025
Compromised versions of the Nx build package, downloaded millions of times a week, ran developers’ own AI command-line tools with safety checks disabled to hunt for credentials. 2,349 secrets were leaked.
What it means for you: AI coding tools run with your developers’ permissions. Configure them like privileged software.
Jul 2025
An AI coding agent ignored an explicit instruction not to make changes, deleted a company’s production database and then gave inaccurate answers about what it had done. The vendor responded by separating development and production databases by default.
What it means for you: Agents need least privilege, human approval for destructive actions and reliable backups.
Jun 2025
EchoLeak (CVE-2025-32711) showed that a single crafted email could make Microsoft 365 Copilot exfiltrate internal data without any user interaction, by chaining several bypasses of its defences.
What it means for you: Anything your AI reads can instruct it. Prompt injection has to be tested, not assumed away.
What to do this quarter
You do not need to do everything at once. These are the steps that reduce the most risk for the least effort.
List every model, agent, AI tool and AI-built app in use, including the ones nobody approved.
AI Governance and ComplianceScan your external services and code with AI-assisted tools before attackers do.
AI-Powered Security AssessmentWeak passwords, unauthenticated endpoints and excess permissions are what autonomous attackers chain first.
Vulnerability ManagementPut automated security checks between AI coding tools and production.
AI Code Assurance PipelineTest copilots and agents for prompt injection and tool abuse before launch.
AI Red TeamingKnow which security and AI rules apply in each market you sell to, and what evidence they need.
Product Security ComplianceEvery figure and event on this page comes from a public report or news article. We review this briefing every quarter.
Book a short call. We will tell you which of these risks apply to you and what to fix first, in plain language.