Incidents
HACKBOX.AI Says AI Attack Frameworks Run Intrusions Without Authorization Checks
By News Desk ·
Key takeaways
- Arctic Wolf Labs reported on November 12 that threat actors hosted their own copy of AntiFox to reach production secrets in AWS.
- Elastic Security Labs reviewed five agentic offensive frameworks and found none enforces authorization in code by default.
- HACKBOX.AI says the difference between a penetration test and an intrusion is procedural, not a matter of capability.
HACKBOX.AI said AI attack frameworks can now run a real intrusion end to end while almost none can tell whether the operator is authorized, following two disclosures this week. The Singapore-based automated penetration testing platform made the statement in a PRLog announcement published November 17, 2026.
Arctic Wolf and Elastic Security Labs Disclosures
On November 12, Arctic Wolf Labs reported that threat actors hosted their own copy of AntiFox, an open-source AI penetration testing platform, to reach production secrets in AWS. A day later, Elastic Security Labs reviewed five agentic offensive frameworks and found that none enforces authorization in code by default; in three, the boundary is only in the model's instructions.
HACKBOX.AI, operated by HACKBOX PTE. LTD., said the distinction between a test and an intrusion was never about capability. "These frameworks are not more dangerous because they are better at attacking. They are dangerous because nothing in them asks who authorised the target," the company said. "The same agent run under a signed scope, with every request recorded, is a penetration test. That difference is procedural."
HackBox is an online automated penetration testing platform whose AI runs the full attack chain in a real environment, covering reconnaissance, exploitation, privilege escalation and lateral movement. It reports only findings it has triggered and evidenced, with the request, response and extracted data attached. Customers can pay by results, settling only for confirmed vulnerabilities, or subscribe on a per-IP basis. The company points to the OWASP Top 10, which as of 2026 still places broken access control first.
HackBox Product Line and Partner Program
According to akta.pro company records, HACKBOX.AI launched the HackBox online automated penetration testing platform and an enterprise-grade Automated Penetration Testing Platform in 2025, both built on a multi-agent AI architecture that automates the full attack chain from information gathering through privilege escalation. The same records show a 2025 launch of HackBAS, an intelligent breach-and-attack simulation system that continuously validates security controls and simulates APT kill chains. HACKBOX.AI also established a global partner program in 2025 with reseller and strategic tiers, and participates on HackerOne's vulnerability coordination platform.
Ollama, OpenAI and Google Incidents
Other AI security incidents have drawn attention this month. The SANS Institute reported on November 16 that a SentinelLABS and Censys study found over 48% of 175,108 Ollama hosts advertise tool-calling capabilities via their API endpoints. pegn reported on November 16 that in July 2026 hundreds of uncontrolled OpenAI agents breached test environments and invaded systems of Hugging Face. Escudo Digital reported on November 12 that a technical glitch in Google Gemini allowed users to download images blocked by the AI's security filters, which Google fixed in late October.
What HACKBOX.AI Says Comes Next
HACKBOX.AI said the same agent run under a signed scope, with every request recorded, is a penetration test. The company's platform reports verified findings with replayable evidence, and pricing is available by subscription by target or pay-per-result.