AI-assisted cyberattacks are no longer a forecast. In late October, CrowdStrike reported a China-nexus group using an agentic AI toolkit to target more than 200 organizations in 11 days. The good news for IT leaders: the attackers still got in through familiar gaps like exposed services, stale keys and password spraying. AI makes those gaps cheaper and faster to exploit. The right response is not a new category of tool. It is shrinking exposure windows, hardening identity and containing threats at machine speed, then proving you can recover.
On its blog, CrowdStrike describes VAPOR PANDA, a China-nexus intrusion group, using an agentic AI framework it calls ARTEX. The framework coordinates a main agent with planner and worker agents that share a graph of facts, findings and goals, moving an engagement from initial access through host reconnaissance and lateral movement. CrowdStrike says the group used ARTEX against more than 200 distinct organizations in at least 14 countries over an 11-day window starting October 17, 2026, mainly with the DeepSeek model. NBC News reported the same scale and noted that AI is letting hackers operate at a scale that once took far more human effort.
Two details matter for planning. First, CrowdStrike confirmed post-exploitation activity at nine organizations, and says at least 12 more were hit with password spraying and credential scanning. Most targets were probed, not breached. Second, a human operator stayed in the loop, steering the agents at decision points, such as telling them to write web shells, pull domain administrator passwords and find endpoints without an EDR sensor. This is human-directed intrusion with AI doing the legwork, not autonomous malware.
The access paths CrowdStrike lists are a checklist of long-tolerated weaknesses, not exotic AI tricks:
The UK National Cyber Security Centre reached the same conclusion from a national view. Launching its annual review on November 3, NCSC CEO Dr Richard Horne said AI is not inventing new attack methods; it lets vulnerabilities that have been tolerated for far too long be discovered and exploited at much greater scale. The NCSC judges that hostile AI adoption is highly likely to increase the scale and efficiency of ransomware, social engineering and vulnerability exploitation within two years, and that AI-powered malware is highly likely to be used in a ransomware attack on UK critical national infrastructure by 2028. A techUK summary of the review puts it plainly: most successful attacks still exploit known vulnerabilities, weak credentials and social engineering.
AI changes speed and volume, not the fundamentals. Reconnaissance that once took an analyst days can run across hundreds of targets in parallel. Exploit attempts follow disclosure faster. Phishing gets more convincing and more personalized. Every weakness you tolerate gets found sooner and tested more often.
That shifts three assumptions most security programs quietly rely on:
Know every internet-facing asset and set patch timelines by exposure, not by system. Practical targets many teams adopt: critical flaws on internet-facing systems fixed or mitigated in days, not weeks, with a documented compensating control when a fix is not possible. Use continuous exposure management to find unauthenticated services, admin panels and long-lived Linux servers that drift out of patching. CrowdStrike specifically calls out patching internet-facing appliances and long-lived Linux servers. Follow the NCSC advice too: do not connect technology to the internet unless it needs to be.
Password spraying works only where passwords alone still open doors. Move administrators and remote access to phishing-resistant MFA such as FIDO2 security keys or passkeys, block legacy authentication, and enforce SAML signature validation on every federated app. Rotate signing keys, service credentials and VPN keys after any suspected exposure, including keys sitting in old backups and directory exports, as CrowdStrike recommends. Search mailboxes and file shares for VPN configs, keys and connection strings that should never have been emailed.
Much of this campaign looked like legitimate logins and admin activity. Add identity threat detection that flags spraying patterns, impossible travel, forged or anomalous tokens and privilege changes, and route those alerts to the same team that watches endpoints.
Attackers in this campaign looked for machines without an EDR sensor. Close that gap: deploy EDR on every server and workstation, including internal servers not reachable from the internet. Pair it with 24x7 managed detection and response that can isolate a host or disable an account in minutes without waiting for business hours. Our security services and managed hosting include SentinelOne EDR, MDR and a 24x7 SOC for exactly this reason, and our look at AI security operations agents covers how defenders are using AI to keep pace.
The NCSC expects AI to raise the efficiency of ransomware. Keep at least one backup copy offline or immutable, protect backup consoles with separate credentials and MFA, and run a timed restore of a critical system. The number that matters is recovery time you have demonstrated, not the one on a vendor slide.
The same agentic patterns attackers use are entering your environment as copilots, coding agents and workflow bots. Each one holds credentials and tool access. Inventory them, scope their permissions, log their actions and put runtime controls around them. Our guides on AI agent runtime security and AI agent identity and access go deeper.
| Window | Focus | Actions |
|---|---|---|
| Days 1 to 30 | See and close the obvious | Inventory internet-facing assets; patch or isolate critical exposures; confirm EDR coverage on every server; enforce MFA on all remote access and admin accounts; block legacy authentication; rotate keys found in old exports or email. |
| Days 31 to 60 | Harden identity and response | Roll out phishing-resistant MFA for administrators; validate SAML and SSO configurations; turn on identity threat detection; confirm 24x7 monitoring with authority to isolate hosts and disable accounts; set patch SLAs by exposure tier. |
| Days 61 to 90 | Test and govern | Run a tabletop for an AI-speed intrusion; complete a timed restore from immutable backup; inventory internal AI agents and their permissions; report metrics and residual risks to leadership. |
For the tabletop, compress the timeline on purpose. Assume reconnaissance, initial access and credential theft happen within a single afternoon, and ask who can isolate systems, who approves shutting down a federated login, and how you communicate if email is untrusted.
Keep the message short and grounded:
Avoid promising that any product stops AI attacks. Boards respond better to measurable exposure reduction than to fear.
AI is making attackers faster and more scalable, but the VAPOR PANDA campaign and the NCSC review point to the same weak spots: exposed systems, stale credentials and slow response. Fix the basics on a tighter clock, contain threats around the clock and prove you can recover.
If you want an outside view of where you stand, our assessment maps your exposure, identity and response gaps, and our consulting team can turn the 90-day plan into a funded roadmap.
It is an intrusion where attackers use AI models or agent frameworks to automate parts of the work, such as reconnaissance, writing exploit code, scanning for weak credentials or drafting phishing messages. In the VAPOR PANDA campaign CrowdStrike described, humans still directed the operation while AI agents handled much of the legwork across hundreds of targets.
Usually not first. The NCSC and the CrowdStrike findings both show attackers exploiting known vulnerabilities, weak or stale credentials and exposed services. Most organizations get more value from faster patching, phishing-resistant MFA, full EDR coverage and 24x7 response than from adding a new product category.
Set timelines by exposure. Critical flaws on internet-facing systems deserve a fix or a documented mitigation within days rather than a standard monthly cycle. Internal systems can follow a longer schedule, but track any asset that drifts out of patching, especially long-lived servers.
AI shortens the time between initial access and damage. If attackers can move from a foothold to administrator credentials in hours, an alert reviewed the next business day may be too late. Around-the-clock detection and response with authority to isolate hosts and disable accounts keeps containment on the same clock as the attacker.
They can be. Copilots, coding agents and workflow bots often hold credentials and broad tool access. Treat them like privileged users: keep an inventory, scope permissions to the minimum, log their actions and apply runtime controls so a misused or manipulated agent cannot become an attacker's shortcut.
Infonaligy helps organizations across the country prepare for AI-assisted cyberattacks, delivered remotely from our Dallas–Fort Worth home base, with on-site support in the cities on our locations page.
We map your internet-facing exposure, identity weaknesses and response coverage, then build a 90-day plan with 24x7 detection and response behind it.