DEV Community

Cover image for CyberStrikeAI recorded 221 executions without requiring approval each time — no violation was proven
Hacks.gr
Hacks.gr

Posted on Originally published at en.hacks.gr Fully Autonomous

CyberStrikeAI recorded 221 executions without requiring approval each time — no violation was proven

Huntback’s review of an exposed CyberStrikeAI workspace covers chat records, execution logs and application source code.

The open-source framework connects a language model to tools for security testing.

The logs show 221 tool calls, but “completed” means a tool finished—not that its requested goal succeeded.

A stored setting showed pre-action approval disabled; researchers could not identify the operator or establish whether the tests were authorised.


The operator was still giving instructions to the program via chat. However, this is not the same as approving each action before it is executed. At 09:48:15 UTC, the recorded setting showed that the approval requirement was disabled.

Huntback did not verify who changed the setting and found no evidence that the program bypassed the check on its own. The four recorded action rejections came from an automatic check that did not receive a response from the model. None were human decisions.

Later, the operator asked the program to list possible actions that would write or delete data and approved the suggestions in the chat. This did not re-enable the setting that would require approval before execution.

A total of 240 tool calls were recorded. 221 were for write operations, while the remaining 19 were for local operations, such as updating drawings and reading files. Of the 221 executions, 209 completed and 12 failed. The indication “completed” means that the tool finished its operation, not that it achieved its purpose. Some completed actions returned authentication or application errors.

↳

READ ALSO

HACKS.GR / RELATED

AI Unlocked

Armornet accepts teams to test platform for continuous monitoring of risks from suppliers

↗

The available results show that a compute service returned information about nodes and tasks. They do not prove that the program gained the ability to execute commands on a computer, accessed an internal network, accessed a data storage system, or made destructive changes to a database.

A search result with at least 10.000 matches does not prove that 10.000 records were downloaded. Also, one program reported that it collected 260 node records, but it did not check for duplicates, and the researchers did not retrieve the entire list.

CyberStrikeAI recorded nine findings, but that doesn't prove that there were nine separate and confirmed security vulnerabilities. Nor does the creation of a new digital wallet prove that login credentials were stolen.

Huntback examined the available files without running any recovered programs, repeating requests, or contacting the services being tested. It did not verify who operated CyberStrikeAI or whether the tests were licensed.

The case suggests that actions were taken without requiring prior approval, not that the program succeeded in a broader attack. Huntback recommended that security teams check whether shared storage or backups expose chat logs, documents, and logs, even when the main application requires a login.

>Stay connected with Hacks.gr

Set us as a preferred source on Google and find us on social media.

Preferred source

↗

end of article


Read the original English article on Hacks.gr

Top comments (0)