Black October: An expert explains how AI is changing cybersecurity

Last week marked a watershed in the information security space. Major incidents occurred within a few days of each other, leading cybersecurity practitioners to dub the month “Black October”.

On October 17, a researcher reported weak private keys used for eMudhra certificates, which could allow hackers to impersonate legitimate websites.

On October 21, the PHP Foundation announced two critical-level vulnerabilities dubbed PHP2Shell. These exploited memory safety, and could allow a remote user to control any PHP server.

The same day, a version of LiteLLM, an open-source repository of LLM providers, was hacked and quarantined, all within a few hours. Separately, GitLab released fixes for a flaw of the highest possible severity, which could allow hackers to remotely control systems.

The intriguing aspect of these incidents is not only their proximity in time, but the role that artificial intelligence (AI) played in them.

AI-assisted discovery and ramifications in PHP2Shell

PHP2Shell presents a great case study in how vulnerability discovery, inordinate amplification, and insufficient support infrastructure formed a highly explosive combination.

Story continues below this ad

PHP is a language that, informally, allows browsers to interact with web servers. Two cybersecurity groups discovered PHP2Shell. AI firm Calif used frontier OpenAI models as part of the latter’s “Patch the Planet” initiative, while Google’s Wiz used in-house models. Both alerted the PHP Foundation, which issued an advisory asking website maintainers to apply patches that it would soon issue for critical vulnerabilities.

Interestingly, the possibility of such exploits had been raised as long ago as 2004. However, researchers could not reproduce a working attack. The precise attacks were only discovered now using AI-assisted techniques.

If you log in to a website or make a purchase, chances are, you are using PHP. Around 70% of all websites reportedly run some version of PHP, and so PHP2Shell could affect innumerable websites.

Unfortunately, implementation details of the exploit were leaked well before a patch was issued. The same day, OpenAI posted that they had found critical flaws in PHP, broadcasting the incident far beyond the cybersecurity community. This triggered a race against time: the PHP Foundation needed to release a fix, developers needed to apply it, and attackers could target several hundreds of millions of vulnerable systems in the meantime. However, much of the patching work was being undertaken by volunteers like Czech programmer Ondřej Surý, and this extra conspicuous bug meant that coming up with a patch fast enough got infinitely harder.

🇨🇿 Hats off to Czech tech expert Ondřej Surý! 👏
As critical PHP vulnerabilities put websites worldwide at risk, Ondřej helped provide security updates for older PHP versions, protecting systems that still rely on them.
An important reminder that the people working behind the scenes can make a real difference to our digital world. 🌐

— Czech Embassy Dublin (@CZinDublin) October 22, 2026

Even after the patch was released, it did not cover older versions of PHP, complicating matters further. Cloudflare’s CEO claimed attempted attacks rose from thousands before the leak to millions eventually. In such a volatile environment, we should revisit responsible disclosure and the actions allowed therein.

To what extent can AI guarantee security?

Story continues below this ad

AI is clearly making it easier to find flaws, as evidenced by PHP2Shell being brought to light after two decades of dormancy.

On October 22, as part of their Cyber Verification Program, Anthropic’s Frontier Red Team announced having checked “over 100 real-world open-source software libraries”. The group said it had used formal verification to discover “bugs that had eluded detection for years, despite trillions of test runs”.

In formal verification, one can cast systems and expected requirements in an abstract, mathematical formulation, and prove whether the system satisfies the requirements. Unfortunately, the requirement for mathematical expertise, as well as scaling issues, hindered its widespread deployment. If AI could overcome these obstacles thanks to its unparalleled access to compute, that could help us show that systems behave as expected or find flaws.

However, this process is difficult to fully automate because of the “intent capture” problem – how does one test whether something has been correctly formalised, without human developers to assess the translation from code to abstraction? Further, since any abstraction necessarily throws away information, resultant proofs are only as good (or as bad!) as the assumptions made for the abstraction. A proof of correctness, therefore, does not mean freedom from all possible security flaws.

Story continues below this ad

AI enthusiasts are going all in, though; Jason D Clinton, an Anthropic engineer, posted on X that “things are going to get a lot worse before they get better”.

as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better. security leaders, never let a crisis go to waste: please go to your boards and ask for help right now. not for ai, for the basics. (we need to hold on for ~18 months.)

— Jason D. Clinton 🔸 (@JasonDClinton) October 22, 2026

Who benefits? And quis custodiet ipsos custodes?

It is not inconceivable that we could be facing a situation not unlike when the fire service operated for profit under the Roman Emperor Nero. Owners of edifices that caught fire could choose to pay to suppress it or sell the ruins afterwards, forcing them to agree to specified terms with the service providers. That Rome eventually burned down, whether or not Nero played a particular stringed instrument during, serves as a cautionary tale we might be doomed to repeat, unless we learn from history.

AI companies are in a similar position today – and seem to already be hinting at it, with Clinton’s post also urging infosec practitioners to “ask for help […] not for ai, for the basics” (sic) – to demand payment for protection against cyber threats that miscreants may unearth using GenAI. This creates a problematic incentive: AI firms can essentially play war profiteer, and sell both attack and defence mechanisms to different parties.

Many of them are planning to go public soon, but are also racking up debt. Anthropic, for example, reported a net loss of over $80 billion in the first half of 2026, largely due to non-cash accounting charges. Offering cybersecurity services could therefore open up new avenues of revenue.

Story continues below this ad

GenAI also requires significant data and compute, and setting up data centres which enable this is starting to directly clash with environmental considerations: Google has recently come under fire for illegally razing 1,300 acres of Finnish forest for its data centres, according to The Guardian. Most claims about AI’s technical abilities, whether in mathematics or cybersecurity, have come in large dumps (722 manuscripts, over 100 systems, etc.) that require a lot of voluntary, thankless human effort to sift through, the lack of which often leads to these being taken at face value without explicit verification.

These concerns must be addressed before we can envisage a fully AI-enabled cybersecurity landscape. For now, we trust in human expertise, like that of Ondřej Surý, to keep our systems running as safely as they can.

Vaishnavi Sundararajan is an Assistant Professor at IIT Delhi, working at the intersection of formal verification and cybersecurity.