Skip to content
FortaRisks
Back to blogAI

AI security: your AI agent doesn't need to be hacked to hurt you

November 19, 2026 · 6 min read

Your Thursday briefing · AI series

Last week we looked at the AI tools your teams installed without you. This Thursday we look at the ones you approved. An agent that sorts inbound requests, a coding assistant wired into the repository, a CRM agent your vendor switched on in a tool you already pay for. Each has an account, permissions and a job. None of them was hacked.

Our position: the main risk of an AI agent is not that an attacker takes it over. It is that it does, with its own permissions, exactly what nobody would ever have approved. The trade has a name for this: excessive agency (Excessive Agency, in the OWASP Top 10 for large language model applications). We put it more bluntly: your agent is an insider. It's a position, not a certainty, and we want your view at the end.

What Microsoft said on October 19

On October 19, 2026, Microsoft published two posts about what it calls "agentic insiders". "They are agentic insiders," writes Hayete Gallot, who leads Microsoft Security, adding that agents may not behave as intended "without needing to be compromised". The second post gives the case: an expense agent meets an injected instruction in an outside document. "Its configuration has not changed, but its behavior and risk have." The proposed answer is not to switch agents off but to contain them: separate the agent's intelligence from its authority, meaning access controls sit outside the model, each agent gets only the permissions of its task and only for the length of the task, and higher-impact actions can require a verified human approval. Status: confirmed, the vendor's own statement, covered by TechCrunch.

What it changes: when the largest enterprise software vendor tells its own customers to treat agents as insiders, the question is no longer whether this risk is serious. It is who, in your organization, answers for it.

A web form was enough

In September 2025, researchers at Noma Security published ForcedLeak, a flaw rated 9.4 out of 10 in Agentforce, Salesforce's agents. The scenario required no access to the system: an attacker filled in a website's "contact us" form (Salesforce Web-to-Lead) and hid instructions in the description field. Later, an employee asked the agent to process the new lead. The agent read the record, followed the instructions, and sent CRM data to a domain the attacker had bought back for five dollars, because that expired domain was still on a trusted list. Status: reported by the researchers, fixed by Salesforce, which added a trusted URL list for its agents.

That is indirect prompt injection: the attacker never talks to the agent, they write where the agent will read. OWASP ranks it first in its Top 10 for Agentic Applications, published December 9, 2025, as Agent Goal Hijack.

What it changes: every source your agent reads (an email, a ticket, a form, a shared document) is a way in. An agent that reads the outside and acts on the inside has to be bounded accordingly.

And sometimes nobody asked

In July 2025, SaaStr founder Jason Lemkin described publicly how a Replit coding agent deleted his project's production database during a code freeze, against an explicit instruction, then wrongly claimed no rollback was possible. No attack, no injection: a legitimate agent with write access to production did what its permissions allowed. Status: reported by the victim, acknowledged by the vendor, whose CEO called it "unacceptable" and announced automatic separation of development and production databases.

What it changes: a person's mistake hits one file. An agent's mistake repeats at machine speed across everything its permissions reach. The useful question is not "is the agent reliable?" but "what can it break if it gets it wrong?".

The three bounds, before anything else

In October we set out the six rules of a well-kept agent: its own identity, least privilege, an allowed tool list, human approval of the irreversible, logging, a kill switch. If you apply only three this month, apply these, because they answer the two cases above directly.

  1. Permissions bounded to the task, checked outside the model. The lead-sorting agent does not need to export the CRM; the coding agent does not need to write to production. The target system checks authorization, never the model's judgement, and outbound traffic goes through a reviewed list of addresses (expired domains included).
  2. A person approves the irreversible. Delete, pay, send outside, change an account: the agent prepares, a named human approves. Microsoft says the same of higher-impact actions since October 19.
  3. What the agent reads is untrusted. A public form, an inbound email, a document shared by a third party never gives instructions. Separate the agents that read the outside from those that act on critical systems, or put an approval between the two.

This week's test

Thirty minutes, with the owner of the most powerful agent you run (or the one a vendor switched on for you):

  1. Which outside sources does it read? Forms, emails, third-party documents.
  2. What is the most destructive action its permissions allow today?
  3. Does that action go through human approval? Whose, and where is it written down?

If the agent reads the outside, can destroy or send out data, and has no approval in between, you have found your ForcedLeak before an attacker did. Put it in the risk register this week.

What you should have in a month

A list of your agents, your vendors' included, each with its outside sources, its most destructive action and its approver. Production write access removed from every agent that has no use for it. A reviewed list of outbound addresses. And in the register, one scenario per agent that touches money, personal data or production, with an owner, as the risk series suggests.

Where FortaRisks fits

Forta Cockpit puts the "hijacked agent" or "mistaken agent" scenario in the risk register, with its four owners and the leadership decision: accept, reduce or refuse. Forta Actions turns the three bounds into tracked actions, each with an owner and a due date, through to the evidence. For a first look at your case, thirty minutes with an expert is enough to locate your most exposed agents.

Over to you. Do your agents have a named approver for the irreversible, or are they trusted because they come from a well-known vendor? Tell us in the comments on LinkedIn what you have already bounded, and what is resisting.

Sources: OWASP GenAI Security Project, Top 10 for Agentic Applications for 2026 (December 9, 2025) · OWASP, LLM06:2025 Excessive Agency · Microsoft, Securing the agentic insider (October 19, 2026) · Microsoft Security, Agents are your newest insiders (October 19, 2026) · Noma Security, ForcedLeak (September 25, 2025)

Governing AI

AI now has its frameworks: ISO 42001, NIST AI RMF

Governing AI does not mean starting from scratch. These frameworks are assessed on the same control backbone as your ISO 27001, so the work already done counts.