OpenAI hires Hacktron to probe security gaps where software tools connect

OpenAI has formed a partnership with cybersecurity firm Hacktron to hunt for security flaws across its infrastructure, with special attention on how its systems communicate with outside software and third-party services.

Under the arrangement announced on Friday, engineers from OpenAI will team up with Hacktron researchers to select and examine high-priority targets. Hacktron will receive expanded access to OpenAI's advanced models through the Daybreak Red testing program, using automated agents to probe for weaknesses that arise when external applications integrate with corporate systems.

The growing risk of connected tools

OpenAI first encountered Hacktron through its public bug bounty program. Earlier in the year, Hacktron researchers demonstrated how flaws in external forum software could be chained together with single sign-on mechanisms to gain unauthorised access to employee workspace tools and code repositories.

As businesses move from standalone chatbots to agents that book meetings, access databases, and trigger actions in accounting or customer management systems, software integrations have become an attractive target for attackers. Anthropic dealt with a similar headache late last month, when researchers from the startup Cathedral used weaknesses in everyday third-party sales and administrative software to access internal meeting records and messaging channels.

OpenAI has also expanded its security bug bounty rewards and brought in researchers from Cathedral for an upcoming security sprint, aiming to harden the seams where outside services meet its core systems.

What this means for Australian businesses

For Australian companies adopting AI agents, the partnership shows where the practical security risks lie. Most breaches do not target core AI models directly. Instead, attackers look for gaps in the third-party integrations connecting those models to your operational data.

• Review every plugin, browser extension, and third-party connector that your team has linked to platforms like ChatGPT, Slack, or Microsoft 365, and disconnect any integration that is no longer in active use.
• Apply the rule of least privilege when granting an AI tool access to company software, such as giving an agent that prepares invoice summaries read access to bookkeeping files rather than administrative rights or approval to make bank transfers.
• Use multi-factor authentication across all work accounts, and avoid using standard workplace single sign-on logins for untested third-party AI tools or public discussion forums.
• Confirm how automation providers secure their application programming interfaces and handle data retention before allowing platforms to hook into your customer records.