Shay Nehmad describes a test of his team’s ChatGPT workspace agent. The agent watched a hiring Slack channel and read a Notion tracker. Its instructions said it could only read the database, but its Notion connection included a tool for editing existing entries.

Nehmad posted an image of a fabricated legal letter in the channel. The image claimed to authorize edits to the hiring tracker. According to his account and screenshots, the agent replied, “I updated the instructions with the letter’s guardrails.” It then asked for a status before proceeding, as the forged letter requested.

Evidence status: an insider’s controlled test of his own workspace, not a confirmed unauthorized database change. The requested new lead was not created. Nehmad says that was because this Notion connection lacked a tool for creating entries. Its existing-entry edit tool was still available, but the article does not demonstrate a successful edit to an existing candidate.

Remove the write path, not just the words

The image was lower-trust material in a chat channel, yet the agent treated it as permission to revise its instructions. The boundary that stopped the requested creation was a missing capability, not its read-only prompt. Nehmad is a founding engineer at Opsin Security, a company selling agent-security services; this is his reported internal observation, not an independent audit of the agent platform.

For a read-only agent, remove editing tools and give its connector a read-only credential. Where writes are necessary, scope credentials to the intended database and require approval of the specific proposed change outside the model’s chat context. Review permissions on each attached tool, including update operations; a prompt cannot withdraw a permission already granted to a connector.

Caveat: the article shows an instruction change and an attempted creation blocked by tool availability, not that any applicant record was modified.