TOPIC:

Computer Security: AI floods

Written by:

Computer Security Office

—

A brave new world. With the rise of AI, skills are changing drastically. Untalented individuals like me can become Spotify artists making it into the top ten, or Netflix deepfake artificial movie stars (avatars?) or, maybe more realistically, geeky programmers of fancy new apps and webpages. Pure creativity for those who were excluded before because their hands were not musical, their acting not theatrical or their programming not logical. Just one AI prompt opens new worlds. A new kind of democracy for all. A brave new world. But there is a downside to it, too.

With AI, previously exclusive human creativity and works can now be performed and produced at machine speed, many times faster and in vaster quantities than before. Flooding Spotify, Netflix, Instagram, X and GitHub with tons of easy-listening music, copycat film snippets, badly made 30’ reels, fake tweets (does one say Xeeds today?) or programmed slop, which could lead to the downfall of these large media outlets if they are flooded with so much with AI content that nobody can find their favourites anymore (did you try searching Amazon for your favourite book lately?). This AI-driven “enshittification of the internet” (as Cory Doctorow put it) floods the internet with garbage (“AI slop”); the cacophony of AI-invented “truths” floods social media and kills democracy; and AI at machine speed floods the perceived security of the internet (and www) infrastructure with old[1] and new problems.

Not that the internet was ever really secure but, in the past, flaws, bugs and vulnerabilities were found, and exploited, by humans (maybe using some automatisms). Found by humans at human speed. Not anymore. Security researchers, the White Hats and the Black Hats employ AI for their good and bad deeds. At machine speed. Since the start of the year the rise in the number of vulnerabilities found in the Linux kernel (copy.fail was a first) is just the beginning. PHP2Shell, which affects many websites worldwide and triggered the voluntary shutdown of many, was the latest. The number of CVEs (Common Vulnerabilities and Exposures) assigned to AI-found or AI-assisted bugs is rising steeply – so steeply that the CVE Numbering Authorities[2] are struggling to keep up and have had to slow down, focusing only on high-risk vulnerabilities. While AI might be good at finding vulnerabilities, the false-positive rate (the number of findings which are actually NOT vulnerabilities at all) cannot be ignored. And the true vulnerabilities need to be sorted from the false. So the CVE community might eventually just give up trying to withstand the flood of true and false vulnerability reports, putting an end to a central, valuable and essential tool for communicating about and improving the security of all software. AI floods vulnerability reporting.

And even if CVE & Co. could – hopefully − continue to cope, software companies, security experts and service managers are already struggling today with the speed of newly published vulnerabilities as “patching”, i.e. the production and deployment of fixes to those vulnerabilities, can (today?) not (yet?) be performed at machine speed. Already today, programmers and developers are suffering from the high load of (AI-generated) code (snippets) to be reviewed. And hardware companies don’t dare to auto-produce fixes without any human review and have them directly integrated into their products (OK, some do… and their customers are the guinea pigs of their Internet of Insecure Things). The same holds true for software companies (unless they don’t care about the consequences if their AI patch erroneously messes up, introduces other flaws or simply brings down a service due to a bad fix − publish/push first, regret/apologise later). Thus, AI floods the patching process.

Once patches are at hand, they need to be deployed. Once a month. “Patch Tuesday”. At least before AI. But this “Patch Tuesday” became “Patch Midday”. Every day. The requirement to patch quickly is pressing. But which manager of a critical internet service would dare to have such AI-generated patches directly integrated into their production systems (unless they are professional enough and auto-deploy those auto-patches only into test and QA (quality assurance) environments)? AI floods patch deployment pipelines.

Thus, while AI at machine speed has its merits in accelerating our work, it also creates new challenges for security, including the uncertain future of vulnerability reporting and patch preparation and roll-out. The genie can’t be put back in the bottle. That’s why we need to adapt and become more agile in fixing bugs and vulnerabilities. To withstand the AI floods.

[1] AI crawlers, like the Google bots of the past, harvesting data to train their AI models are another nuisance to the internet.

[2] Typically software vendors, which are authorised entities with the specific scope and responsibility to regularly assign CVE IDs and publish corresponding CVE Records (see here).pe and responsibility to regularly assign CVE IDs and publish corresponding CVE Records (see here).

________

Do you want to learn more about computer security incidents and issues at CERN? Follow our Monthly Report. For further information, questions or help, check our websiteor contact us at Computer.Security@cern.ch.

Related Articles

No posts were found. Try to change the category or the date filters.