For security, AI is a land of contrasts

Welcome to Runtime! Today: Cybersecurity professionals hit the desert to discuss the pros and cons of AI's impact, OpenAI drops GPT-5, and the latest enterprise moves.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Summer camp for hackers

Each year in early August thousands of cybersecurity professionals — everyone from government officials to slick product salespeople to mischievous hackers — descend on Las Vegas for the Black Hat and DefCon conferences. Black Hat, which wraps up Thursday, is the more staid affair focused on the state of enterprise cybersecurity, and like everything else in enterprise tech these days conference speakers spent a lot of time this week discussing the impact of AI.

As Amazon CSO Steve Schmidt told Runtime in June, right now enterprise security teams and vendors who have employed generative AI as defenders have the upper hand on the elements that would like to use the same technology to launch attacks. This week Microsoft and Google showed off the results of their work using generative AI to fight malware and plug vulnerabilities.

• Microsoft's Project Ire was designed to "reverse engineer" suspected malware and determine its threat level without having to consult the security team, with the goal of scaling "the system’s speed and accuracy so that it can correctly classify files from any source, even on first encounter."
• However, as The Register pointed out, the system only detected 26% of all the malware presented to it in a test, although it did a pretty good job reducing the number of false positives that Chainguard CEO Dan Lorenc told Runtime earlier this year have plagued early AI security tools.
• Google also showed off a new tool called Big Sleep that detected vulnerabilities in 20 open-source projects without human intervention.
• But as Ars Technica's Dan Goodin noted, Google declined to comment on the number of false positives generated by its tool.

On the offensive side, voice-phishing attacks (which we are apparently calling "vishing") were in focus this week after Google disclosed on Tuesday that customer data was stolen from one of its Salesforce databases earlier this year using that attack vector.

• "Deepfake vishing – fraudulent phone calls that leverage AI‑generated voice clones – has rapidly evolved into one of today’s most sophisticated social‑engineering threats," Group-IB wrote in a detailed report released this week describing the threat and possible countermeasures.
• The group behind the recent attacks "has demonstrated repeated success in breaching networks by having its operators impersonate IT support personnel in convincing telephone-based social engineering engagements," Google said in a blog post earlier this year describing the technique.
• Cisco also revealed this week that it had been hit with a similar attack targeting its Salesforce data, and to be clear no one is attributing any blame for these breaches to Salesforce.
• This was exactly the scenario that some security professionals worried about in the months after the launch of ChatGPT, and Visa announced a new advisory practice designed to help its customers learn more about this type of threat.

Cybersecurity has always been an arms race; attackers force defenders to step up their game, which then sets off a search for new attack surfaces. Generative AI is a new challenge, given that natural-language commands can be used to obtain sensitive data without having to learn how to exploit a buffer overflow.

• “Any sufficiently capable language model is a monster from Lovecraftian horror,” Cprime's Aaron Bockelie told CSO.
• But legendary cybersecurity expert Mikko Hyppönen confirmed Schmidt's assessment of the current state of AI security in his keynote at Black Hat Wednesday.
• “AI is the key” to mounting a defense, he said, according to Tech Republic. “It’s the biggest technological revolution I’ve seen in my life.”
• Still, if AI agents really do transform the next generation of business software they'll present an enormous new attack surface to protect, and dozens of companies rolled out software and services this week ahead of that potential showdown.

Five on it

OpenAI dropped GPT-5 Thursday, almost two and a half years after the debut of GPT-4 and nearly a year after it was originally expected to arrive. "I think having something like GPT-5 would be pretty much unimaginable at any previous time in human history," said CEO Sam Altman, according to the BBC, and it's hard to argue that the Mayans would have been a little bewildered by the launch video.

GPT-5 simplifies the naming conventions across OpenAI's roster of models, and also features a "real-time router" that decides whether to use a "smart, efficient model" or a "deeper reasoning model" depending on the nature of the prompt. Benchmarks cited in its release blog post compare it favorably to basically every other model ever released, but that's what any benchmark cited in any launch release is going to show.

Early real-world reviews were positive: "It’s still an LLM — it’s not a dramatic departure from what we’ve had before — but it rarely screws up and generally feels competent or occasionally impressive at the kinds of things I like to use models for," said Simon Willison, a widely followed AI expert, after having access to the new model for several weeks. He also noted that OpenAI's pricing for the new model was "aggressively competitive," which should entice developers to give it a whirl.

Enterprise moves

Anand Subbaraman is the new CEO of Icertis, taking over for Samir Bodas (who will remain executive chairman) after serving as COO at the contract-management company for more than a year.

Richard Wong is the new CFO of Fastly, joining the edge computing company after similar roles at Benchling and Houzz.

Diana Kelley is the new CISO at Noma Security, following security leadership roles at Protect AI, Microsoft, and IBM.

The Runtime roundup

AWS will give U.S. government agencies an aggressive discount on cloud services over the next several years, one month after Oracle agreed to a similar deal.

The city council of Tucson, Arizona, voted unanimously Wednesday to block a new data-center project with links to AWS, underscoring how difficult building new data centers in the U.S. is about to get over the next few years.

"Microsoft is cautiously onboarding Grok 4 following Hitler concerns," according to The Verge, and folks, what a time it is to be alive.

Thanks for reading — see you Tuesday!