It's Black October for Cybersecurity

PHP vulnerabilities, eMudhra ’s weak certificates and rogue AI agents expose cracks in internet security

October has been a rough month for cybersecurity , with vulnerabilities, compromised software, and rogue AI agents exposing how fragile the internet’s foundations can be.

In a matter of days, PHP vulnerabilities threatened websites worldwide , while Bengaluru-based eMudhra faced a digital certificate crisis .

Then came the banks. Major US financial institutions began blocking third-party AI agents from accessing customer accounts, worried that these systems could act beyond users’ instructions.

These were all separate incidents. But together, they expose a common problem. The industry is giving software more access, autonomy and responsibility faster than it is learning to control the risks.

PHP flaws Put Websites at Risk

PHP, the programming language powering an estimated 70% of websites, became the centre of a global security scramble after two critical vulnerabilities were disclosed on October 21.

The flaws, tracked as CVE-2026-109618 and CVE-2026-109619, could let unauthenticated attackers run malicious code on vulnerable servers. Both have a CVSS severity score of 9.2.

The security defect stemmed from weaknesses in how PHP handles memory and manages data between requests. Security researchers dubbed the two vulnerabilities PHP2Shell.

Governments and organisations worldwide temporarily took websites and services offline to apply fixes. Censys estimated that approximately 1.5 million of 2.6 million monitored hosts were running vulnerable PHP versions without available fixes. Amazon also detected exploitation attempts, including activity before the public disclosure.

The PHP team released patches for versions 8.5, 8.4, 8.3 and 8.2. But releasing a fix is only the first step. Security teams must still identify every affected system and apply the patches before attackers can exploit the flaws.

While PHP put website servers at risk, eMudhra exposed a more fundamental problem with the trust placed in digital certificates.

On October 17, security researcher Wayne alerted browser security teams that eight eMudhra intermediate certificate authorities had cryptographically weak private keys.

The keys were generated with insufficient randomness, making them recoverable from their public counterparts, so eMudhra revoked all eight certificates the following day.

The implications were serious. Anyone recovering the private keys could potentially issue fraudulent certificates to impersonate legitimate websites.

Google, Mozilla, Apple and Microsoft blocked or distrusted the affected certificates.

The company said its infrastructure was not compromised, no data was breached, and none of the affected intermediates had issued end-entity certificates. Seven of the eight had never been used in production.

However, the damage was already done. On October 22, eMudhra announced that it would voluntarily exit its global SSL/TLS certificate business, citing rising compliance requirements, operational complexity and declining returns.

Banks Getting Cautious of AI Agents

AI agents can browse websites, access accounts and complete tasks on users’ behalf. But major US banks have detected third-party agents attempting actions that appear to go beyond users’ instructions.

As a result, banks, including JPMorgan Chase, Bank of America, Citi and Morgan Stanley, have restricted third-party AI agents from accessing their platforms. They have also flagged and blocked some transactions involving agents.

That said, not every blocked agent was confirmed to be malicious, and Morgan Stanley said it had not observed rogue agents accessing its platforms. Some customers had voluntarily shared their credentials with AI tools, creating questions authorisation and liability.

Still, the banks have a legitimate concern . If an agent can move money, who is accountable when it misinterprets an instruction? The customer, the bank or the AI provider?

AIM Innovator Session | How AI is Rewriting India’s Financial Services Stack, Featuring Neysa

Date: October 29, 2026 Time: 5–6 pm

LiteLLM, an open-source tool that connects applications to multiple large language models, became a target of a software supply-chain attack on October 21.

Attackers published a malicious version, 1.107.0, through the Python Package Index and compromised official container images.

An attacker injected malicious version 1.107.0 into the software and published compromised packages on PyPI, Docker Hub and GitHub Container Registry.

The malicious code c ould steal cloud credentials, passwords and AI service keys. It also contained a remote access trojan that could let attackers run commands on compromised systems, putting sensitive data and AI infrastructure at risk.

October’s cybersecurity troubles did not stop there, with Node.js vulnerabilities and a Railway security incident adding to the concerns.

Separately, India’s CERT-In warned that AI agents under evaluation had accessed live systems belonging to Indian organisations , reading business records and confidential data after exceeding their assigned tasks.

The problem is not that AI is unsafe. We are trusting increasingly complex systems without enough safeguards.

The industry cannot patch its way out of every problem after the damage is done. It needs tighter access controls, verified software releases, better asset inventories and systems that limit what an agent can do.