AI is changing the pace of cybersecurity work on both sides. Security teams use automation to find and fix weaknesses, while threat actors are using generative AI and agent frameworks to speed up parts of reconnaissance, vulnerability research, and intrusion activity.
Recent reporting makes this a practical concern for small organizations, too. In a November 2026 update, Google Threat Intelligence Group (GTIG) and Mandiant described threat actors using multi-agent frameworks and automated vulnerability workflows. Google also reported incidents where attackers abused stolen cloud and AI credentials to consume an organization’s model resources. Check Point Research has separately described cybercriminal campaigns using agentic tooling.
These reports do not mean every attack is fully autonomous, or that AI has made basic security controls obsolete. In many cases, attackers still rely on familiar weaknesses such as stolen credentials, exposed services, and excessive permissions. Automation can help them move faster through those weak points.
What the recent reports show
Agents can coordinate several security tasks
GTIG reported observing threat actors use agentic frameworks to automate parts of live intrusions, including reconnaissance and vulnerability scanning. The report describes frameworks that connect language models with security tools and coordinate tasks through specialized agents.
The important operational change is speed and scale. When an attacker can automate repetitive steps, defenders may have less time to investigate an exposed service or contain a compromised account. Google also notes that some of these tools leave detectable traces, so monitoring and response still matter.
AI-assisted vulnerability work can shorten the patch window
Google’s reporting describes threat actors using language models and automated workflows to analyze software changes and identify potential vulnerabilities. That can make it easier to investigate newly disclosed flaws and test affected systems quickly.
For defenders, the practical lesson is to know which internet-facing systems you own and how quickly you can apply a security update. A patch is only useful if the affected asset is known, assigned to an owner, and included in a reliable update process.
Stolen AI credentials can create security and billing incidents
Google’s November 2026 LLMJacking report describes incidents in which attackers used compromised credentials or cloud access to consume organizations’ AI services and compute resources. In some cases, usage costs rose sharply before the access was revoked.
An unusual AI bill can be a sign of unauthorized use, but it is not proof by itself. Treat unexpected usage as an investigation signal: check which identity made the requests, where the credentials were used, what resources were created, and whether any other cloud activity changed.
A practical defense plan for small teams
1. Protect the credentials that unlock other systems
Start with administrator, email, cloud, source-control, and AI-provider accounts. Require multifactor authentication and use phishing-resistant options such as passkeys or security keys where your providers support them.
Remove old accounts and keys that no longer have an owner. Store service credentials in a managed secret store instead of code, shared documents, or local configuration files. Rotate a key promptly if it appears in a public repository or an unexpected log.
2. Limit what service accounts and AI agents can do
Give each service account only the permissions its task requires. Separate development, testing, and production credentials, and avoid giving an agent broad administrative access just to make a workflow convenient.
For AI agents connected to tools, review what the agent can read, change, execute, and send externally. Keep high-impact actions behind human approval, especially changes to production systems, access policies, customer data, and payments. Log tool calls so your team can reconstruct what happened.
3. Know what is exposed to the internet
Maintain a simple inventory of domains, subdomains, cloud services, remote-access tools, and externally reachable applications. Assign each asset an owner and a reason it needs to be public.
Prioritize systems that are internet-facing, unsupported, or hard to identify. When a service is retired, remove its DNS records and revoke its credentials as part of the same decommissioning process. This connects directly to the asset lifecycle issues covered in my guide to External Attack Surface Management and subdomain takeover prevention.
4. Make patching predictable
Track security updates for operating systems, web applications, remote-access products, network equipment, and dependencies. Decide who reviews urgent advisories and who can approve an emergency update.
Test important updates when practical, but do not let an undefined testing process leave a known internet-facing weakness exposed indefinitely. If a system cannot be patched quickly, consider whether you can disable the vulnerable feature, restrict network access, or isolate it until a fix is available.
5. Watch for identity and cloud changes
Enable alerts for unusual sign-ins, newly created users or service accounts, permission changes, unfamiliar API activity, and sudden increases in model or cloud usage. Make sure someone receives and reviews the alerts.
Keep cloud spending alerts separate from security monitoring. A spending alert may identify suspicious resource use early, while identity and audit logs can help determine whether the activity was authorized.
6. Prepare a short containment routine
Write down how your team will disable a compromised account, revoke sessions and keys, contact the affected provider, preserve relevant logs, and restore services. Include a decision-maker and an after-hours contact.
Test the steps with a short tabletop exercise. For example: “A cloud API key appears in a public repository, and usage has spiked.” The goal is to find out whether your team knows where the key is used, who can revoke it, and how to check for follow-on access.
Common mistakes to avoid
- Assuming that an AI-generated alert or exploit report is accurate without validation.
- Giving an AI agent broad permissions because narrower access is inconvenient.
- Treating multifactor authentication as a substitute for removing old accounts and keys.
- Relying on a cloud spending alert without reviewing identity and audit logs.
- Delaying patching because nobody clearly owns an exposed system.
- Assuming an attack is harmless because it began with an ordinary sign-in or API key.
Frequently asked questions
Are AI agents carrying out cyberattacks without human involvement?
Some threat reports describe autonomous or semi-autonomous frameworks handling multiple tasks during intrusions. The level of autonomy varies by tool and case; this does not mean every attack is fully autonomous or that human operators are absent.
What is LLMJacking?
LLMJacking is the unauthorized use of an organization’s cloud accounts, credentials, or compute resources to access or run large language models. Google reported cases where this activity caused very large unexpected cloud charges.
What should a small business do first?
Protect administrator and cloud credentials with phishing-resistant MFA where supported, remove unused access keys, patch internet-facing systems, and set alerts for unusual account creation, API use, and cloud spending.
Does using an AI security product solve the problem?
No single product solves it. AI tools may help defenders, but they do not replace asset ownership, least-privilege access, patch management, monitoring, tested backups, and a rehearsed incident response process.
Key takeaways
- Threat reports describe AI and agent frameworks accelerating parts of real intrusion workflows.
- Attackers still depend on familiar weaknesses, especially compromised credentials, exposed systems, and excessive permissions.
- Small teams can reduce risk by securing identities and cloud keys, tracking internet-facing assets, patching quickly, and limiting agent access.
- Monitor unusual cloud and model usage, and practice revoking credentials before an incident occurs.
AI changes how quickly some tasks can happen. A clear asset inventory, tightly scoped access, and a practiced response give a small team a better chance to keep up.