Artificial intelligence has exploded in popularity in recent months, to the point that some big tech companies believe it’s time to establish a common set of standards for building and deploying these new technologies. is.
Google on Friday announced the Secure AI Framework (SAIF), a conceptual framework for securing AI systems.
SAIF develops an ecosystem that keeps pace with AI developments, expands detection and response with AI in mind, integrates automation into AI, and red-teams with models developed for AI. We aim to carry out exercises.
Google will work closely with government standards bodies to help develop the NIST AI Risk Management Framework and the industry’s first AI certification standard, the ISO/IEC 42001 AI Management Systems standard.
In concrete action, Google said it plans to expand its bug bounty program to encourage industry research on AI safety and security. We also plan to publish several open-source tools to help practice the SAIF elements for AI security.
“The research community plays an important role in the AI ecosystem, and we are proud to already have such a relationship with security researchers,” said Phil Venables, Chief Information Security Officer, Google Cloud. said. “Last year, we paid out more than $12 million in bounties to security researchers who tested our products for vulnerabilities. We’ve been working with the Yoko community, and we also have a research arm, Google DeepMind, that works on these issues.”
Most security professionals thought it was a good thing that a big company like Google took such a strong step in promoting SAIF, but an area most security professionals are learning as they go along. Some believe that many challenges remain.
“We are just starting to think about this and are making analogies with existing cybersecurity disciplines,” said John Vanbeneck, chief theater hunter at Netenrich.
Bambeneck said it makes sense to have a bug bounty program if you’re talking about software applications, but AI doesn’t even really know what penetration testing actually looks like. pointed out.
“We’re actually making things up on the fly, so we just need to fix it and see what happens,” Bambeneck said. “In that sense, getting some out there is a good first step, because at least it gives the industry a starting point to figure out what works and what doesn’t.”
SAIF is off to a great start, based on several principles found in the NIST and ISO frameworks, said Sounil Yu, chief information security officer at JupiterOne. Today, the industry needs a bridge between current security controls and those specifically required for AI systems.
“The main difference with AI systems that makes SAIF particularly attractive and necessary is that AI systems don’t have many opportunities to make mistakes,” says Yu. “AI safety is a very important principle to consider in the early stages of the design and development of AI systems, as it can have catastrophic and irreversible consequences. Incorporating safety principles early will help AI systems better align with human values and reduce potential abuse of these technologies. can be prevented.”
Piyush Pandey, Chief Information Officer at Passlock, said that similar types of controls were needed in the same way that Sarbanes-Oxley Act (SOX) created the need for segregation of duties (SOD) controls for financial processes. He pointed out that it is clear that AI is also necessary. system.
Pandey said SOX requirements have quickly been applied to the business applications that run these processes, resulting in control testing becoming its own industry, along with software solutions, audit firms and consulting firms, helping customers ensure control effectiveness and compliance. I said I was helping to prove it.
“For SAIF to be right, we need to define controls to give organizations a starting point to make their AI systems and processes safer,” said Pandey.
