OpenAI is working to enhance the ability to combat cyber attacks for its Atlas AI browser. However, they also admit that prompt injection attacks, a type of attack that manipulates AI agents to execute malicious instructions often hidden in websites or emails, are a risk that will not disappear in the short term - this raises questions about the safety of AI agents operating on the open internet.
malware attacks, like online scams and social techniques, are unlikely to be completely resolved, OpenAI wrote in a recent blog post, details how the company is enhancing Atlass defense capabilities against relentless attacks. The company admits that the perfect mode in ChatGPT Atlas expands the security threat surface.
In addition to OpenAI, the UK National Cyber Security Service (NCSC) warned earlier this month that reminder-based encrypted attacks targeting generative AI applications may never be completely downplayed, leaving websites at risk of falling victim to data leaks.
The UK government advises cybersecurity experts to reduce the risk and impact of script-based attacks, rather than thinking that attacks can be intercepted.
On OpenAI's side, the company said: "We consider instant vaccination a long-term AI security challenge and we will need to continuously strengthen our defense against it."
What is the company's solution for this impossible task? A quick, proactive response cycle that the company says is showing initial positive signs of helping detect new internal attack strategies before they are exploited "in practice".
OpenAI stressed that immediate and highly effective anti-malware attack is difficult, but they are relying on large-scale testing and a faster patch-up cycle to enhance system security before they appear in real-life attacks.
An OpenAI spokesperson declined to reveal whether Atlas's security update will significantly reduce the number of successful malware attacks, but said the company has partnered with third parties to enhance Atlas's ability to combat rapid malware attacks before its launch.
