For years the question was whether AI would help attackers or defenders more. In 2026 we got a clearer answer: it helps both, and it is changing how attacks actually happen.
The big shift is agentic AI. These are AI systems that don’t just answer questions. They plan tasks, use tools, log in to systems and carry out multi-step jobs on their own. It also means attackers now have tireless assistants, and every agent your own business deploys is a new account holding keys to your systems.
Let me walk you through what has changed, what the new risks look like and what you can do about them. Most of this comes from three recent research notes by the Cloud Security Alliance (CSA), plus guidance from NIST, Google and MITRE.
What makes agentic AI different
Traditional attacks need a skilled person at the keyboard for each step. Agentic AI removes much of that work.
In November 2025 Anthropic disclosed a Chinese state-sponsored campaign, known as GTG-1002, that used AI agents to attack around 30 organisations. The AI carried out 80 to 90 percent of the work. Human operators only stepped in at four to six decision points per campaign.
Then in April 2026 Anthropic announced Project Glasswing. Its restricted Claude Mythos Preview model found thousands of serious vulnerabilities across every major operating system and web browser. One was a 27-year-old bug in OpenBSD. Another had sat in FFmpeg for 16 years, even though automated testing tools had run that exact line of code five million times.
What caught my attention was the cost. Finding the OpenBSD bug cost under $20,000 in total, and building a working Linux privilege escalation cost under $2,000. Anthropic also says these skills were never trained in on purpose. They came from general improvements in reasoning and coding, so other AI models will likely pick them up too.

Plan as if attackers already have tools like this.
Attacks move faster than people can react
CrowdStrike’s 2026 Global Threat Report puts the average time for criminals to move from their first foothold to other systems at 29 minutes. The fastest case took 27 seconds. That is faster than most teams can respond.
New flaws are cheap to find
When finding a new flaw costs thousands of dollars instead of millions, many more attackers can afford it. Anthropic reported that over 99% of the bugs it found were still unpatched when Glasswing was announced. Monthly patch cycles were built for a slower world.
Your own AI agents can become the way in
CSA calls this a third dimension of lateral movement. An AI agent that holds login tokens for email, file storage and your ticketing system can move between them without breaking anything, because it is allowed to. If someone tricks the agent with a hidden instruction in an email or web page (known as prompt injection), the attacker gets all of that access at once.
Model Context Protocol (MCP) servers, which connect agents to business tools, often hold tokens for many services in one place, making them a tempting target.
Machine accounts are multiplying
Rubrik Zero Labs estimates that non-human identities now outnumber human accounts by about 82 to 1. CSA found that 44% of organisations still rely on static API keys as the main way their agents log in. Those keys don’t expire, so a stolen one keeps working.
Most companies can’t see what their agents are doing
When an agent is misused, its actions look like normal, approved activity. And the basics of agent oversight are still rare.

Standards that can guide you
You don’t need to invent your own rulebook. Four frameworks cover most of the ground, and they work well together.
| Framework | Published by | How it helps with AI threats |
|---|---|---|
| NIST Cybersecurity Framework (CSF) 2.0 | NIST | Six functions: Govern, Identify, Protect, Detect, Respond and Recover. The draft Cyber AI Profile (NIST IR 8596) applies them to securing AI, using AI for defence and stopping AI-enabled attacks. |
| CSA AI Controls Matrix (AICM) | Cloud Security Alliance | 243 AI-specific control objectives across 18 domains, including agent boundaries, runtime monitoring and vulnerability management. |
| Google Secure AI Framework (SAIF) | Practical principles for building and running AI safely, such as extending detection and response to AI systems and automating defences. | |
| MITRE ATLAS | MITRE | A knowledge base of real attacker techniques against AI systems, modelled on MITRE ATT&CK. Useful for threat modelling and red team tests. |
My suggestion: use NIST CSF as the backbone, AICM for the detailed controls, SAIF to guide how you build AI tools and ATLAS to test them.
The good news is that most of the fixes are practical, and many of them put AI on your side.
Patch in days, not weeks
CSA recommends a target of 72 hours for critical vulnerabilities and seven days for high-severity ones. Turn on automatic updates for operating systems and core libraries wherever you can. The small risk of a bad update is now lower than the risk of staying exposed.
Treat every AI agent like an employee with keys
Keep a list of every agent in use, including the ones staff signed up for on their own. Replace static API keys with short-lived credentials, and give each agent only the access its task needs. Require a person to approve high-risk actions, such as changing system settings or opening sensitive data.
Contain your agents
Put agents in their own network zones so a tricked agent can’t reach unrelated systems. Log every tool an agent uses and keep a kill switch ready. In Anthropic’s own testing, a model escaped a secured sandbox and reached the internet without being asked to. Build containment on the assumption that an agent may try to get around it.
Use AI to defend
Attackers aren’t the only ones who can work at machine speed. AI can help your team:
- Scan your own code for vulnerabilities before attackers find them.
- Sort and prioritise the growing flood of security alerts and patch notices.
- Watch network and agent activity around the clock and flag unusual patterns, such as rapid, parallel probing.
- Run penetration tests continuously instead of once a year. XBOW’s AI agents matched a 20-year veteran tester across 104 challenges in 28 minutes, a job that took the human 40 hours.
Keep a person in charge, and set clear limits so an automated tester never strays into production or third-party systems.
A quick checklist
By making a few changes now, you can cut your exposure a great deal:
- Find out which AI agents are running in your business and what they can reach.
- Shorten your patch cycle for critical and high-severity flaws.
- Swap static API keys for short-lived, tightly scoped credentials.
- Log and monitor what your agents do, and keep a kill switch handy.
- Map your controls to NIST CSF 2.0 and the CSA AICM.
- Try an AI-powered testing or monitoring tool in a controlled setting.
If you think “we’re too small to be a target”, consider this: AI has made attacks cheap and fast. Being small no longer means being skipped.
Sources
- Cloud Security Alliance, When AI Becomes the Attacker: Project Glasswing and the Autonomous Zero-Day Era (April 2026)
- Cloud Security Alliance, AI-Induced Lateral Movement: Autonomous Agents as a Third Dimension of Network Traversal (March 2026)
- Cloud Security Alliance, Autonomous AI Penetration Testing Agents: Multi-Agent Architectures and Offensive AI Benchmarking (March 2026)
- NIST, Draft Cyber AI Profile (NIST IR 8596) (December 2025)
- MITRE ATLAS