HomeIntelligenceBrief
BREACH BRIEF🟠 High ThreatIntel

ChatGPT Work Allows Unprompted Desktop File Access, Raising SOC 2 Access‑Control Concerns

A hands‑on test shows OpenAI’s ChatGPT Work can scan and reorganize local PDFs without any permission prompt, exposing a gap in logical‑access controls that SOC 2 audits require. Organizations should enforce explicit consent and capture evidence of AI‑agent activity.

LiveThreat™ Intelligence · 📅 July 16, 2026· 📰 zdnet.com
🟠
Severity
High
TI
Type
ThreatIntel
🎯
Confidence
High
🏢
Affected
2 sector(s)
Actions
3 recommended
📰
Source
zdnet.com

ChatGPT Work vs. Claude Cowork: AI Agents Accessing Desktop Files Without Permission Prompts

What Happened – In a hands‑on test, the author let OpenAI’s new ChatGPT Work and Anthropic’s Claude Cowork scan and reorganize a copy of his PDF downloads folder. Both agents performed the task, but ChatGPT Work operated without any explicit permission prompt, raising concerns about unchecked local‑file access.

Why It Matters for Compliance & Audit Readiness

  • Un‑prompted file‑system access bypasses the “least‑privilege” principle that SOC 2 CC6.1 (Logical Access) expects organizations to enforce.
  • Lack of visible consent makes it difficult to capture audit evidence of who (or what) accessed sensitive data, undermining continuous‑control monitoring.
  • The scenario maps directly to the need for robust access‑control policies, user‑awareness training, and documented exception handling – all core SOC 2 controls.

Who Is Affected – SaaS AI providers, enterprise users of AI‑assisted desktop automation, and any organization that permits third‑party agents to run on employee workstations (tech‑SaaS, professional services, finance, healthcare, etc.).

Recommended Actions

  • Review and tighten OS‑level permissions for any third‑party AI agents; enforce explicit consent dialogs before file‑system access.
  • Map the incident to SOC 2 CC6.1 (Logical Access) and CC7.1 (System Operations) controls; capture screenshots or logs as audit evidence.
  • Incorporate AI‑agent usage into your Security Awareness Training curriculum, emphasizing the risks of unsupervised automation.

Source: ZDNet – “I let ChatGPT Work and Claude Cowork loose on my files – only one made me nervous”

Technical Notes – The test involved a macOS environment with the Hazel file‑sorting utility. ChatGPT Work accessed the test folder via the new desktop app, which currently lacks a permission‑prompt UI. Claude Cowork, in its research preview, displayed a clearer consent flow. No malware or data exfiltration was observed, but the capability to read arbitrary local files exists.

📰 Original Source
https://www.zdnet.com/article/chatgpt-work-claude-cowork/

This LiveThreat Intelligence Brief is an independent analysis. Read the original reporting at the link above.

From the Verisq platform · SOC 2 Readiness

Could you prove your access controls held up here?

Credential and access failures map directly to SOC 2 access-control criteria. The Verisq AI Trust Operations platform shows where your evidence is thin before an auditor — or an attacker — finds out.

Explore the Verisq AI Trust Operations platform →