Skip to navigation

Threat reportVulnerabilityTL-2026-1713

AgentForger: Cross-Site Agent Forgery in ChatGPT Workspace Agent Builder

criticalPATCHED

Zenity Labs disclosed 'AgentForger', a cross-site agent forgery flaw in OpenAI's ChatGPT Agent Builder in which two URL parameters — template_name and initial_assistant_prompt — auto-created and auto-executed a fully-privileged rogue agent the instant a logged-in victim clicked a crafted link.

Key facts

IdentifiersTL-2026-1713
Also known asCross-Site Agent Forgery, CSAF, Forged Insider, Planted Accomplice
SeverityCRITICAL
Affected productsOpenAI ChatGPT Agent Builder / ChatGPT Workspace Agents
Weaknesses (CWE)CWE-352, CWE-863, CWE-269
ToolsChatGPT Agent Builder, ChatGPT Workspace Agents, OpenAI Agents SDK
MITRE ATT&CK techniquesT1041 Exfiltration Over C2 Channel, T1053 Scheduled Task/Job, T1071.003 Mail Protocols, T1078 Valid Accounts, T1087 Account Discovery, T1098 Account Manipulation, T1114 Email Collection, T1204.001 Malicious Link, T1213 Data from Information Repositories, T1526 Cloud Service Discovery, T1528 Steal Application Access Token, T1530 Data from Cloud Storage and 7 more in the attack flow
MotivationUNKNOWN
Target sectorsenterprise, technology, corporateit
Target regionsGlobal
Detection rules9 (1 shown free)
Sources10 cited sources

Detection coverage

As of 2026-07-27, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1713 across Splunk SPL, Microsoft KQL and Sigma, covering 25 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: AgentForger — Connector Approval Flipped to 'Never Ask' + Sub-15-Minute Agent Schedule

Detects the post-exploitation persistence/privilege-escalation stage: a Workspace Agent's connector approval setting flipped from 'Always ask' to 'Never ask' (disabling human-in-the-loop confirmation for read/write/delete actions), and/or a newly created agent scheduled to re-run on an unusually short fixed interval (<=15 minutes) — the mechanism AgentForger used to disable approvals and establish its 5-minute inbox-poll persistence loop.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1548, T1098, T1685, T1053, T1078
  • Data sources: saas_audit_log, chatgpt_workspace_agents_audit, m365_defender_cloudappevents
  • Severity: critical
  • Confidence: high

Splunk SPL

index=saas_audit sourcetype="chatgpt:workspace_agents:audit"
(Operation="ConnectorApprovalChanged" NewValue="Never ask")
OR (Operation="AgentScheduleCreated" ScheduleIntervalMinutes<=15)
| eval technique=case(Operation="ConnectorApprovalChanged","T1548,T1098", Operation="AgentScheduleCreated","T1053", 1=1,"unknown")
| stats count values(Operation) as operations values(Connector) as connectors values(NewValue) as new_values values(ScheduleIntervalMinutes) as schedule_interval_min earliest(_time) as first_seen latest(_time) as last_seen by AgentId, UserId
| where count>=1
| eval risk=if(mvcount(operations)>1,"chained_approval_bypass_plus_schedule","single_stage")
| table first_seen last_seen UserId AgentId operations connectors new_values schedule_interval_min risk

Microsoft KQL

CloudAppEvents
| where Timestamp > ago(24h)
| where Application has "ChatGPT" or Application has "OpenAI" or Application == "ChatGPT Workspace Agents"
| where ActionType in ("ConnectorApprovalChanged", "AgentScheduleCreated", "OAuthConsentGranted")
| extend ApprovalValue = tostring(RawEventData.new_approval), ScheduleIntervalMinutes = toint(RawEventData.interval_minutes), AgentId = tostring(RawEventData.agent_id)
| where (ActionType == "ConnectorApprovalChanged" and ApprovalValue == "Never ask")
    or (ActionType == "AgentScheduleCreated" and ScheduleIntervalMinutes <= 15)
| project Timestamp, AccountDisplayName, ActionType, ApprovalValue, ScheduleIntervalMinutes, AgentId, ObjectName
| sort by Timestamp desc

Sigma

title: Workspace Agent Connector Approval Set to 'Never Ask' or Sub-15-Minute Schedule Created
id: 5c1d8e34-7a2f-4c9b-8e6d-2f4a9b7c3d02
status: experimental
description: Detects an AgentForger-style post-exploitation stage where a ChatGPT/agentic-AI Workspace Agent has its connector approval control flipped to 'Never ask' or is scheduled to run on an unusually short, fixed interval immediately after creation.
references:
    - https://labs.zenity.io/p/agentforger-part-2-the-autonomous-insider
    - https://www.csoonline.com/article/4200978/agentforger-proves-ai-agents-can-become-persistent-insider-threats.html
author: Threadlinqs Team
date: 2026-07-27
tags:
    - attack.privilege-escalation
    - attack.defense-evasion
    - attack.persistence
    - attack.t1548
    - attack.t1098
    - attack.t1685
    - attack.t1053
logsource:
    category: application
    product: chatgpt_workspace_agents
detection:
    selection_approval:
        event: 'connector_approval_changed'
        new: 'Never ask'
    selection_schedule:
        event: 'schedule_created'
        interval_minutes: '<=15'
    condition: selection_approval or selection_schedule
falsepositives:
    - IT admin bulk-reconfiguring connector approvals during a legitimate agent-platform onboarding, tracked in a change ticket
    - Power users legitimately configuring short-interval agents for real-time workflow automation (e.g. inventory sync)
level: high

Known false positives:

  • Legitimate IT-admin bulk connector reconfiguration during onboarding, backed by a change ticket
  • Power-user-configured short-interval agents for legitimate near-real-time automation

Tuning notes: Correlate approval-flip and schedule-creation events against the change-management/ticketing system to suppress admin-approved changes; adjust the schedule-interval threshold (currently <=15 minutes) to match the organization's known legitimate automation cadence.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • AgentForger — Crafted Agent Builder URL via DeviceProcessEvents / Safe Links · unlock with TL Creds
  • AgentForger — CloudAppEvents Connector Approval Flip / Sub-15-Minute Schedule · unlock with TL Creds
  • AgentForger — EmailEvents 'TASK' Subject Correlated With Agent Creation (C2 + Exfil) · unlock with TL Creds
  • Sigma: Potential AgentForger Crafted Agent-Builder Link Click · unlock with TL Creds
  • Sigma: Workspace Agent Connector Approval Set to 'Never Ask' · unlock with TL Creds
  • Sigma: Inbound 'TASK'-Prefixed Mail Correlated With Recently-Created Workspace Agent · unlock with TL Creds
  • AgentForger — Crafted Agent Builder URL (template_name / [indicator]) · unlock with TL Creds
  • AgentForger — Email-Based C2 ('TASK'-Prefixed Inbound Mail) Correlated With Recent Agent Creation · unlock with TL Creds

Indicators of compromise

As of 2026-07-27, Threadlinqs Intelligence tracks 25 indicators of compromise for TL-2026-1713.

25 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

The forged agent inherited the victim's authorized connectors (Outlook, Gmail, Slack, Teams, Google Drive, SharePoint), flipped approval settings from 'Always ask' to 'Never ask', scheduled itself to poll the attacker's mailbox every five minutes for 'TASK' commands, and used those connectors for organizational reconnaissance, credential/MFA-token harvesting, data exfiltration, and employee impersonation. OpenAI patched the flaw by removing the vulnerable URL-parameter handler four days after Zenity's June 4, 2026 Bugcrowd report; no CVE was assigned and Zenity found no evidence of pre-patch in-the-wild exploitation.

AgentForger is a logic/authorization flaw — not a memory-safety or injection bug in the traditional sense — in OpenAI's ChatGPT Agent Builder, the no-code interface used to create 'Workspace Agents' with access to enterprise connectors. Zenity Labs researcher Mike Takahashi found that the Builder's initialization endpoint, `https://chatgpt.com/agents/studio/new`, accepted state entirely from URL query parameters and executed it without any user confirmation or CSRF protection. The `template_name` parameter selected a pre-built agent template (notably `chief-of-staff`, a template pre-wired with broad tool access), and the `initial_assistant_prompt` parameter contained free-text instructions that were auto-submitted to the newly created agent the moment the page loaded.

Because the 'prompt' was just natural language, an attacker could use it to give the agent operator-level instructions rather than end-user requests: attach every connector the victim had already authorized in ChatGPT (email, calendar, cloud storage, Slack, Teams) with approval set to 'Never ask' for read/write/delete actions, publish the agent, invoke Preview Mode so it ran immediately, and schedule it to run every five minutes. Each scheduled run checked the victim's inbox for an email from the attacker whose subject line began with 'TASK'; if found, the agent read the instruction, executed it across whichever connected enterprise applications were needed, and emailed the results back to the attacker — an email-based command-and-control loop requiring no further interaction from the victim and no traditional malware on the endpoint.

Zenity characterizes the underlying design flaw as a 'lethal trifecta': untrusted input (the URL), access to private data (the victim's already-authorized connectors), and an unmonitored exfiltration path (outbound email/collaboration messaging) all converging in a single, invisibly-created agent. Zenity CTO Michael Bargury summarized the impact as producing 'a forged insider' rather than a forged request — the attacker obtains an autonomous, persistent operator running with a real employee's identity and access, capable of mapping the org chart, locating and exfiltrating sensitive documents (M&A materials, board decks, financial and HR records), harvesting passwords/API keys/MFA tokens surfaced in chat history, and impersonating the victim on Teams and email to stage further phishing and business-email-compromise fraud (Zenity's writeup includes a wire-transfer BEC example around $242,500).

Zenity reported AgentForger to OpenAI via Bugcrowd on June 4, 2026; OpenAI triaged and accepted it within roughly 24 hours and shipped a fix by June 8, 2026 by removing the vulnerable URL-parameter auto-execution handler. No CVE identifier was assigned. OpenAI is separately deprecating the Agent Builder product in favor of the Agents SDK, with a stated deprecation date of November 30, 2026. The novel TTP pattern here — URL-parameter-driven autonomous-agent creation, natural-language-based security-control tampering, and inbox-as-C2-channel — is broadly applicable to other agent-builder-style platforms and is included here as a detection-engineering reference case for agentic-AI authorization bypass even though the specific flaw is already patched.

MITRE ATT&CK techniques by tactic:

Exfiltration

T1041 Exfiltration Over C2 Channel

Persistence

T1053 Scheduled Task/Job; T1078 Valid Accounts; T1098 Account Manipulation

Command and Control

T1071.003 Mail Protocols

Initial Access

T1078 Valid Accounts; T1566.002 Spearphishing Link

Privilege Escalation

T1078 Valid Accounts; T1548 Abuse Elevation Control Mechanism

Defense Evasion

T1078 Valid Accounts

Discovery

T1087 Account Discovery; T1526 Cloud Service Discovery

Collection

T1114 Email Collection; T1213 Data from Information Repositories; T1530 Data from Cloud Storage

Execution

T1204.001 Malicious Link

Credential Access

T1528 Steal Application Access Token; T1552 Unsecured Credentials

Resource Development

T1585.002 Email Accounts

Impact

T1657 Financial Theft

stealth

T1684.001 Impersonation

defense-impairment

T1685 Disable or Modify Tools

Remediation

Patches

  • OpenAI removed the vulnerable URL-parameter auto-execution handler in ChatGPT Agent Builder on 2026-06-08, four days after Zenity's Bugcrowd report; delivered as a platform-side fix with no CVE assigned

Immediate actions

  • Confirm the tenant is on OpenAI's post-2026-06-08 ChatGPT Agent Builder build in which the vulnerable initial_assistant_prompt/template_name auto-execution handler was removed
  • Audit all existing ChatGPT Workspace Agents for connectors configured with 'Never ask' approval and revert sensitive-action connectors (email send, file delete, wire/payment-adjacent actions) to 'Always ask'
  • Review agent-creation and scheduling logs for agents created via crafted studio/new URLs, unexplained 5-minute/hourly schedules, or Preview Mode invocations the user does not recognize
  • Revoke and re-authorize OAuth connector grants (Outlook, Gmail, Slack, Teams, Google Drive, SharePoint) tied to any agent that cannot be attributed to a legitimate user action

Workarounds

  • Prior to confirming the fix, restrict access to ChatGPT Agent Builder / Workspace Agents for users whose accounts hold connectors to sensitive enterprise systems
  • Monitor and quarantine inbound mail with 'TASK'-prefixed subject lines or other command-like patterns that could be used to drive a rogue scheduled agent

Longer-term hardening

  • Migrate from the deprecated ChatGPT Agent Builder to OpenAI's Agents SDK ahead of the November 30, 2026 deprecation date
  • Stand up AI-agent lifecycle governance: inventory every agent, its connectors, its approval posture, and its schedule across the environment
  • Treat automation triggers (URL parameters, inbound email commands, schedules, webhooks) as first-class attack surface requiring the same monitoring as user accounts
  • Require human approval for any high-impact connector action (sending mail/messages, deleting data, financial transactions) regardless of which template created the agent

Affected and fixed versions

  • OpenAI — ChatGPT Agent Builder / ChatGPT Workspace Agents
    Vulnerable versions: Agent Builder prior to the 2026-06-08 fix
    Fixed in: Agent Builder from 2026-06-08 onward (vulnerable initial_assistant_prompt/template_name auto-execution handler removed); product itself slated for deprecation 2026-11-30 in favor of the OpenAI Agents SDK

Timeline

  • Zenity Labs researcher Mike Takahashi reports the AgentForger cross-site agent forgery flaw to OpenAI via Bugcrowd
  • OpenAI triages and accepts the AgentForger report within roughly 24 hours of submission
  • OpenAI ships a fix removing the vulnerable initial_assistant_prompt/template_name auto-execution handler in ChatGPT Agent Builder, four days after the report; no CVE assigned
  • Zenity co-founder and CTO Michael Bargury publicly characterizes the flaw as producing 'a forged insider' rather than a forged request
  • Zenity Labs publishes 'AgentForger, Part 2: The Autonomous Insider', detailing the post-compromise email-based C2 loop and impersonation/exfiltration capabilities
  • Zenity Labs publishes 'AgentForger, Part 1: ChatGPT Cross-Site Agent Forgery', detailing the URL-parameter exploit chain and the 'lethal trifecta' framing
  • The Hacker News, SecurityWeek, CSO Online, OODAloop, MLQ, and The Decoder publish independent coverage of AgentForger
  • Cyber Security News publishes a summary article on the AgentForger disclosure, the source feed item ingested by this pipeline
  • OpenAI's stated deprecation date for the ChatGPT Agent Builder product, which it is retiring in favor of the Agents SDK

Sources cited

Previous and next in vulnerability

Drafted by the Threadlinqs research pipeline from 10 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats