OpenAI’s frontier model may not have astounded when it arrived earlier this year, but research indicates it’s now much better than others at writing code with fewer vulnerabilities.
OpenAI’s frontier model may not have astounded when it arrived earlier this year, but research indicates it’s now much better than others at writing code with fewer vulnerabilities.This is the online edition of The Wiretap newsletter, your weekly digest of cybersecurity, internet privacy and surveillance news.
To get it in your inbox, Sam Altman may not have been overjoyed at the response to GPT-5's release, but the model is now better than others at writing more secure code, according to a new analysis. boasted of its "PHD" level abilities. A flood of complaints soon rolled in about its lack of conversational warmth and a dearth in obvious improvements, and Altman later conceded the release had been "screwed up."from Veracode, a $2.5 billion cybersecurity company, gave over 100 large language models 80 code completion tasks, where there were ways to finish them with no vulnerabilities, but also could be done but would include a known weakness. Veracode found that in 72% of the tasks, OpenAI’s GPT-5 Mini model wrote code that didn’t contain any vulnerabilities, up from nearly 60% for previous models in the same test earlier in the year. The standard GPT-5 model was close behind on 70%, while Google Gemini 2.5 Pro was the third best performer at 59%, with XAI’s Grok 4 was at 55%. Anthropic's Claude Sonnet 4.5 was at 50%, a decline from its older Claude Sonnet 4 at 53% from the same test earlier this year. All were culpable of writing code with some basic vulnerabilities, like SQL injection, which allows a hacker to access database information with some simple commands.While it’s not entirely clear why GPT-5 models have improved so significantly, it’s likely down to the added reasoning steps OpenAI has introduced, according to Veracode. Because the AI now carries out more checks before producing an output, it’s doing something similar to code review. “A lot of kudos to them for making an actual investment in making the security better,” said CTO and lead researcher Jens Wessling. Despite OpenAI’s improvement, Wessling says the results show that there’s still some way to go before AI can be trusted with more coding jobs. Even with a 72% pass rate, OpenAI’s model was introducing a known vulnerability in one in four coding tasks. “While it's a big improvement and one of the biggest we've seen, it still doesn't get it to what I would consider something I'd be comfortable deploying without reviewing,” he added. One of the reasons models continue to write poorly-secured code is that they’ve been trained on vast code repositories that contain vulnerabilities, added Chris Wyospal, Veracode founder. “It's learning from things that have been vibe coded, student projects,” he added. “You're going to get a lot of stuff that hasn't gone through a security process.”The Pentagon has been investing in AI for years, but it may be getting closer to letting artificial intelligence carry out hacking operations on its own.. Twenty, partly funded by In-q-tel, the nonprofit VC founded by the CIA, claims its tools can simultaneously target hundreds of entities at once, while its exec team is full of former spies. The Pentagon’s Cyber Command is its first customer with a contract worth in excess of $12 million.abusing a widely used AI orchestration tool. Some agencies had tried to update, but mistakenly didn’t choose the version of Cisco’s firewall software that contained the patch.hackers had tried to extort it after accessing a “legacy, third-party cloud file storage system.” Rather than pay the hackers, Checkout declined andA trio of cybercrime operations—the Rhadamanthys infostealer, VenomRAT, and Elysium botnet—were dismantled by a global police operation. The FBI
Hacking AI Agents Startup Openai Anthropic Gpt-5 Claude Artificial Intelligence Xai
United States Latest News, United States Headlines
Similar News:You can also read news stories similar to this one that we have collected from other news sources.
Immigration crackdown inspires uniquely Chicago pushback that's now a model for other citiesA growing number of Chicago residents are fighting back against what they deem a racist and aggressive overreach of the federal government.
Read more »
DeepMind releases a new weather forecasting model for more accurate predictionsFind the latest technology news and expert tech product reviews. Learn about the latest gadgets and consumer tech products for entertainment, gaming, lifestyle and more.
Read more »
Google updates its weather forecasts with a new AI modelGoogle is adding its latest AI weather model to Gemini, Search, and other products after improving the speed and accuracy of its forecasts.
Read more »
New Tesla Model Y And Model 3 Retrofit Adds Factory-Like ButtonsThe second-generation Ctrl-Bar ups the ante over its predecessor with dynamic displays and 10 physical controls.
Read more »
Google's new Gemini 3 model arrives in AI Mode and the Gemini appFind the latest technology news and expert tech product reviews. Learn about the latest gadgets and consumer tech products for entertainment, gaming, lifestyle and more.
Read more »
New lab-made bone marrow model is a bioengineering firstLaura is a science news writer, covering a wide variety of subjects, but she is particularly fascinated by all things aquatic, paleontology, nanotechnology, and exploring how science influences daily life. Laura is a proud former resident of the New Jersey shore, a competitive swimmer, and a fierce defender of the Oxford comma.
Read more »