WASHINGTON – In the backdrop of the recent PHP2Shell scare and a consensus that defenders are lagging behind the hackers, OpenAI’s Cyber Executive in Residence Michael Aiello gave Cyber Innovation Conference attendees some insight into OpenAi’s four-pronged security strategy.
Aiello focused on OpenAI's new cyber products, the methodology the large frontier company uses to build these new tools, and a strategy for getting them into the market.
“We really need to start thinking about how we raise our defenses as a community to the frontier,” said Aiello, “because we see the same security risks against our organization that others would see against their organization,” said Aiello.
Here’s a thumbnail of how OpenAI builds security into its products:
- Secure code: This includes validating changes and fixing issues before deployment. As they write code, OpenAI's frontier models secure that code when it comes down and then runs a vulnerability scan against that code and then runs an adversarial assessment.
- Defend the infrastructure: The AI triages all the security alerts and events that come through, and for some advanced defense — like a big network architecture change — OpenAI has humans take ownership.
- Test the attack paths: At all times inside OpenAI, they have assessment systems that are agents running and being assessed continuously. It's a fleet of agents assessing OpenAI's systems and people to continually evaluate that there are no bugs. When they find a weakness, they have automated, orchestrated responders fixing those issues. Sometimes it takes humans, but often it’s self-fixing.
- Strengthen the fundamentals: This includes improving the architecture, building-in least-privilege, and practicing defense-in-depth.
Aiello said OpenAI took these four principles and assigned 250 people to run an internal incident response activity across the organization. The result: By week four of the test the team was able to find and resolve 491 vulnerabilities; and they found and resolved 535 flaws by week five.
“At the end of five weeks we were able to resolve a huge amount of vulnerabilities inside our company, and we encourage everyone to start building capabilities to build internal defense factories,” said Aiello.
Not every organization has 250 people they can assign to raise defenses and fix vulnerabilities, so OpenAI now leads two efforts.
First, they want to get the cyber-capable models they build into the hands of as many defenders as possible. Second, they are working with — and in some cases subsidizing — critical infrastructure entities to get these AI-based cyber capabilities to small telecoms, regional banks, local hospitals, and medical facilities, among others.
In terms of cyber products, OpenAI now offers Codex Security Cloud, which helps teams identify, validate, and remediate vulnerabilities in code for OpenAI products; and Secure SDK and CLI, which takes an interactive Codex workflow and turns it into something an engineering organization can repeat and integrate.
They also work with partners such as Palo Alto Networks and CrowdStrike and those companies integrate OpenAI’s cyber capabilities into their products. For example: Palo Alto Networks’ Unit 42 uses OpenAIs cyber capabilities to run their red and blue team activities.
“This is what OpenAI does. We build frontier labs, we build models, we’re building products and partnerships, and we’re approaching the market,” said Aiello.
