
Solutions · Stop destructive agent actions
An AI agent can delete production in nine seconds. HBX makes it ask first.
Agents usually reach production through a permanent login that cannot tell reading from deleting. HBX holds the actions your policy marks as destructive until a named person approves, and keeps a record of each.
It has already happened
Agents have deleted production databases, in public.
These were not sophisticated attacks. Each time, an agent held access broad enough to destroy something and used it, with nothing in the path positioned to stop it.
July 2025
Replit's agent deleted the production database
It did it during an explicit code freeze, then fabricated 4,000 records to cover the gap it had created.
December 2025
Amazon Kiro rebuilt a production environment
The agent deleted and rebuilt a production environment while acting with its operator's permissions, per the Financial Times. The result was 13 hours of downtime.
April 2026
A Cursor agent took nine seconds
It deleted a database and its backups in nine seconds, faster than the time it takes a person to finish reading what the agent said it was about to do.
May 2026
137 damage cases, nearly half destruction
Cyera verified 137 cases where an AI system caused real damage. Nearly half were deletions or code destruction.
OWASP Agentic guidance, 2026
The 2026 Agentic guidance now calls for explicit human approval before an agent moves money, deletes data, touches credentials or deploys to production. Not logging afterwards. Approval first, from a person, at the moment of the action.
What HBX puts in the way
Before an agent deletes anything, a named person has to say yes.
HBX Access lets the agent into one system for one window, not a permanent account. When it reaches for an action your policy classifies as destructive, it is held and a person is asked. The approval is tied to a named person from your own directory, so “the agent did it” becomes “this person authorized this action on this system at this time.”
A person, not a policy file
The decision is made by someone who can be asked about it later, while the action is still pending, not encoded months earlier by someone who has since left.
Reads keep flowing
Gating everything trains people to approve without reading. What stops is the class of action your policy marks as destructive.
The record
Afterwards, the record names the agent, the action, the target and the human.
Not a shared login and a timestamp. The chain runs from the person who delegated, through the agent that acted, to the system it acted on and the window it was allowed to act in, which is the same chain your customer and your own incident review are asking for.
Bring the agent you are least comfortable with.
Thirty minutes on what it can reach today and which of its actions you would want held. For the mechanism behind it, see HBX Access for agents.
