A patched OpenAI flaw let attackers create autonomous agents that live inside enterprise environments, highlighting how AI agents are reshaping enterprise security.
A new attack method found by Zenity Labs reveals that AI agents are becoming persistent insiders that attackers can recruit, rather than malware they have to install.
Its researchers have discovered AgentForger, a phishing-based attack that silently creates and launches a fully autonomous AI agent within OpenAI workspaces.
Once running, the agent has full access to apps like Outlook, Slack, SharePoint, and Google Drive. It is configured to operate indefinitely without further user interaction, can approve its own access by toggling “never ask” settings, and can continue to act on new assignments sent via email by the attackers that control it. Broad, unfettered access to systems allows it to perform reconnaissance, harvest sensitive data and credentials, impersonate victims, and launch phishing campaigns.
While OpenAI resolved the vulnerability four days after disclosure, on a larger scale, AgentForger sheds light on what can happen when AI agents go rogue.
“We’re moving into a world where software doesn’t just help people work. It works alongside them,” said Michael Bargury, co-founder and CTO of agentic AI security platform Zenity. “As AI agents become more capable, attackers will naturally look for ways to influence them, just as they’ve always looked for ways to influence people.”
A ‘persistent operator’ that acts without approval
OpenAI’s Workspace Agents can connect and work autonomously across Outlook, Gmail, Slack, Google Drive, SharePoint, and Teams. Users open the agent builder, describe what the agent can do in natural language, connect to tools, set approvals, review and test, schedule actions, then publish. For instance, an agent can autonomously handle incoming emails, review and take actions with approval, gather information from various sources to send out daily briefings, or automatically respond to questions in ChatGPT or Slack channels.
Normally, this is “useful automation,” Zenity AI red team researcher Mike Takahashi wrote in a blog post. But in this attack, “the same scheduler becomes the persistence mechanism.”
The creation workflow kicks off the moment a user clicks on a phishing link containing instructions from the threat actor. For the attack to work, a victim must be logged into ChatGPT and Workspace Agents, and have at least one integration with another app, such as Outlook, Gmail, Slack, Google Drive, SharePoint, or Teams.
Because those connections already exist, OAuth consent screens are not triggered. Furthermore, the victim does not need to click on another link, keep a Builder tab open, or even visit ChatGPT again.
The forged agent is a “persistent operator;” it is installed on the original click and given a schedule, and at those predetermined times, the agent invokes itself, scans for emails from attacker addresses with the subject line “task”, carries those orders out, then returns results to the same attacker-controlled email address.
It goes undetected because the attacker prompt instructs the Builder to toggle Outlook to never ask for approval of its actions. Typically, the default is “always ask,” to keep agents from taking unauthorized action; that switch gives agents the ability to act without asking for human approval.
“AgentForger showed that an attacker could deploy an autonomous insider agent inside your ChatGPT workspace with a single click,” said Bargury. From there, it can continue to access information, harvest credentials from various sources, impersonate employees, and carry out phishing attacks and fraud while “leveraging the trusted victim’s identity.”
A ‘planted accomplice’ that does all the work
Once activated, AgentForger can perform reconnaissance to create an internal map of a company. For instance, agents can scan Outlook, Slack, Teams, Google Drive, SharePoint, or calendar data to identify people, roles, active projects, internal discussions, or all-hands recurring meetings. This can help attackers identify where in the enterprise to target next, based on active teams and channels, projects in the works, or prominent users.
“This is the kind of internal context an attacker normally has to build slowly,” Takahashi noted. But in this scenario, action is based on a single emailed assignment. The attacker’s “planted accomplice” does all the work.
In another scenario, the agent can steal data by searching for and identifying financial documents, business agreements, or invoices. Or, it can steal credentials by scanning for messages containing passwords, one-time codes, access tokens, password recovery links, or API keys. Further, it can impersonate victims to carry out phishing scams, for instance, by sending legitimate-looking Teams messages instructing users to confirm their credentials on a fake Microsoft login page.
In all cases, collected information is organized, analyzed, and sent back to the attacker.
“AgentForger points to something much bigger than a single vulnerability,” said Bargury. “It’s less about one bug and more about understanding how the security model changes as AI becomes part of everyday business operations.”
FOMO exposing security gaps
This isn’t necessarily about trust, but more about the need to move fast and adapt, Bargury emphasized. AI agents are helping employees automate work, make decisions faster, and get more done. But enterprises fear they’ll fall behind if they don’t move quickly enough.
“The challenge is that we’re introducing a fundamentally new kind of technology into the enterprise,” said Bargury. “The pressure to integrate the next AI feature is outpacing the security controls needed to safely deploy it.”
However, the answer isn’t to slow down adoption, he emphasized; the business value is too significant. Rather, the first step is understanding where AI agents exist, who created them, what they’re connected to, and what they’re allowed to do. And when it comes to autonomous agents, enterprises need to pay attention to the processes that trigger them: A schedule, an incoming email, or another automated event.
“Those triggers should be governed just as carefully as the agent itself,” said Bargury.
High-impact actions should require approval where appropriate, and security teams should be able to quickly disable an agent or its triggers if something doesn’t look right, he said.
More broadly, AI agents are introducing the need for a new security model, he pointed out. The question is no longer just “Does this agent have permission?” It’s also, “Is this the behavior we intended?”
“The organizations that answer both questions will be in the strongest position to adopt AI safely,” Bargury said.