CyberStrikeAI ran 221 tools after approval enforcement was turned off
Original title: CyberStrikeAI Raises Security Concerns After 221 Tool Executions Without Human Oversight
Huntback found an exposed session database, not a confirmed intrusion, shell, or data-lake breach.
Huntback Research examined a 6.69 MB SQLite snapshot from an exposed CyberStrikeAI workspace used in a November 16, 2026 offensive testing session. Approval enforcement was off by 09:48:15 UTC, and 221 of 240 tool calls had persistent execution IDs, all from the main orchestrator. The records do not establish a remote shell, internal-network access, stolen funds, or destructive database changes. Researchers recommend correlating reviewer failures, policy changes, and later executions.
Unconfirmed: The records do not identify the operator, who disabled approval, or whether the testing was authorized; the agent's critical ratings were not independently validated.
- The database held one conversation with 16 messages and 240 runtime tool calls.
- An automated reviewer rejected four approval requests; none was a human decision.
- The runtime marked 209 executions completed and 12 failed.
- Artifacts were retrieved from 38.76.169[.]207 on TCP port 8010.
- Ocean Protocol documents compute discovery, so node data alone is not proof of disclosure.
- Products
- CyberStrikeAI
- Organizations
- Huntback ResearchOWASP
From the original article
3 min read at gbhackers.com
An exposed CyberStrikeAI workspace recorded 221 tool executions during a November 16, 2026, offensive testing session in which formal approval enforcement was turned off partway through. The session documents a loss of tool-level oversight, not an autonomous intrusion. The records identify neither the operator nor whether the testing was authorized.
By Mayura Kathir · gbhackers.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.