AI Hacking Framework CyberStrikeAI Runs 221 Tool Executions Without Human Approval

Blog WriterCybersecurity News - Original News Source is cybersecuritynews.com


An exposed CyberStrikeAI workspace recorded 221 tool executions in a single session on November 16, 2026. The logs show no human decision at the framework’s tool approval gate, and the saved configuration later showed that the gate had been switched off.

A person still directed the AI agent through chat. The difference matters: telling an agent what to do is not the same as approving each action before it runs.

CyberStrikeAI is an open-source penetration testing framework, not a newly discovered malware family. It connects a language model to security tools and stores their results.

Earlier reporting on CyberStrikeAI multi-agent activity described a separate case involving 44 agents. In the November 16 session, every recorded tool call came from one main agent, even though the workspace was configured in supervisor mode.

Researchers from huntback identified the exposed workspace and examined its conversation database, diagnostic log and application source.

Their CyberStrikeAI workspace investigation did not establish who operated the framework or whether the testing was authorized. The researchers reviewed the artifacts without running recovered programs, replaying requests or contacting the services under test.

AI Hacking Framework CyberStrikeAI

The runtime logged 240 tool calls, but only 221 had execution IDs that matched entries in the database’s execution table.

The other 19 were local actions, such as updating plans and reading files, that did not receive those IDs. Of the stored executions, 209 were marked completed and 12 failed.

A completed status meant the tool finished running, not that the action it attempted succeeded. Some completed runs returned authentication or application errors.

The database recorded four approval rejections, all issued by an automated reviewer after it failed to get a response from its model. None was a human decision.

At 09:48:15 UTC, the saved configuration showed approval enforcement was off. Huntback could not determine who changed the setting and found no evidence that the AI agent bypassed the control on its own.

The operator later asked the agent to list potentially harmful write and delete actions for review, then approved the proposed items in chat.

That approval did not restore the software control that could stop a tool call before execution. The repeated instructions also distinguish this session from cases covered in the FBI’s ARTEX security warning, where some intrusions reached privileged access in under three hours.

What the CyberStrikeAI Session Actually Proved

The strongest evidence was that a compute service returned node and job data. The retained outputs did not establish that the agent gained a shell, reached an internal network, breached a data lake or made a destructive database change.

A search result showing at least 10,000 matches did not establish that 10,000 records were downloaded. Likewise, a script reported collecting 260 node entries, but it did not check for duplicates, and researchers did not recover the full list.

The agent entered nine findings in the database, but those were not nine independently verified vulnerabilities. A newly generated wallet did not prove that someone else’s credentials were stolen, and a catalogue of available tools did not establish permission to run them.

As coverage of AI attack automation trends shows, agents can drive real tool activity. Investigators still need returned data and system logs to verify what that activity achieved.

For security teams running AI agents, the clearest detection opportunity is the sequence: reviewer errors, a change to approval settings and renewed tool activity.

Huntback also recommends checking whether file servers or backup routes expose conversation databases, uploaded documents or runtime logs, even when the main application requires authentication.

Indicators of compromise (IoCs):-

Type Indicator Source-reported association
IPv4 38.76.169[.]207 Host that served the workspace files
Port TCP 8010 File-serving service
URL path /CyberStrikeAI/data/conversations.db Exposed conversation database
URL path /CyberStrikeAI/log/diagnostic-2026-11-16.log Diagnostic log for the recorded session
Host path /root/CyberStrikeAI/ Application working path
SHA-256 a0b4be0a70e35d0d30809e0856e9860d4265df65e41fef2e93405e166d6edd36 Conversation database snapshot (6,688,768 bytes)
SHA-256 4da5383d7a1bc2a83c20a39510227aad34072876fa47bfb20fd10aeb48b676eb Diagnostic log (6,560 bytes)
Model configuration deepseek-v4-pro Locally recorded model identifier

Note: IP addresses and domains are intentionally Qdefanged (e.g., [.]) to prevent accidental resolution or hyperlinking. Re-fang only within controlled threat intelligence platforms sQuch as MISP, VirusTotal, or your SIEM.

Give your SOC the alert context on the latest threats to stop incidents quickly -> Integrate Threat Intel from 16K SOCs

The post AI Hacking Framework CyberStrikeAI Runs 221 Tool Executions Without Human Approval appeared first on Cyber Security News.