Black October: How to Protect Your Accounts and Online Assets from AI-Driven Attacks

In the third week of October 2026, a cluster of security incidents got a name: Black October. AI-assisted researchers found critical flaws in PHP, a popular AI gateway library shipped credential-stealing malware for a couple of hours, and a certificate authority had to give up its business after its keys were shown to be weak. Security executives told the press to expect more of the same.

If you run accounts, a website, a domain, a crypto wallet or any cloud service, this is a good week to tighten things up. This guide explains what happened in plain terms, how it can reach you, and which defences are worth your time. It is written for ordinary users first, with sections for website owners and developers further down.

What Black October actually was

The name covers several separate events. The facts below come from reporting by the Guardian, CNBC and Business Insider, Wikipedia's summary article, and the vendor posts they cite (linked at the end).

• 17 October: eMudhra intermediate certificates. A researcher reported that private keys for eight intermediate certificate authority certificates run by the Indian provider eMudhra were weak, because they were generated with too little randomness in April 2025. Anyone able to recover those keys could forge certificates that vulnerable clients would trust. Google Chrome and Mozilla Firefox pushed revocations without needing a browser restart. eMudhra said none of the affected authorities had issued end-entity certificates and that it found no breach. On 22 October it announced it would leave the global SSL/TLS certificate business.
• 21 October: PHP2Shell. The PHP team released emergency updates for two critical vulnerabilities (CVE-2026-109618 and CVE-2026-109619) that could let an unauthenticated attacker run code on a vulnerable web server. Wiz researchers demonstrated remote code execution against WordPress, Drupal and Joomla installations, using AI-assisted discovery and exploit development. Patched releases are PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12. CNBC reports that Wiz and the AI firm Calif found the flaws in August using models from OpenAI and Google, and held the details back while a fix was prepared. The US National Security Agency told operators to patch urgently, and the PHP project's message was blunt: if you run PHP, update now.
• 21 October: LiteLLM supply-chain attack. Attackers published a malicious LiteLLM version, 1.107.0, to PyPI and pushed compromised container images through Docker registries. The malware went after browser credentials, cloud secrets, developer tokens and API keys for AI services. By LiteLLM's own timeline, the bad PyPI packages were live for under two hours (11:39 to 13:23 UTC).
• 21 October: GitLab. GitLab fixed a critical remote-code-execution flaw affecting self-managed instances. The Hacker News reported a CVSS score of 10.0.

Government and private sites in the US and UK went offline briefly while operators patched, and the Associated Press reportedly saw similar disruption in Finland and Japan. Taking a site down in that situation is usually a precaution, not proof of compromise.

Cloudflare's CEO Matthew Prince said his company saw thousands of attempts to exploit the PHP flaws in the ten hours before the public announcement, rising to millions soon after. Sam Altman said he no longer thinks all bugs can be patched, and Anthropic's Jason Clinton warned of a "vulnerability tsunami". Former CISA director Chris Krebs told CNBC that 3% of all critical vulnerabilities ever discovered were found in the last month.

One number needs care. Roughly 70% of websites use PHP, according to estimates CNBC cites, and Wiz's headline used that figure. It describes how many sites run the language, not how many were vulnerable. Sites on patched versions, or behind a managed host that updated for them, were not exposed in the same way.

How this reaches your accounts

It helps to be precise here. Nothing above is "AI hacking your Gmail". AI made it cheaper and faster to find deep bugs and to turn them into working exploits. You are exposed in three indirect ways.

1. Services you use get breached. If a site you have an account on runs unpatched PHP, your email, password hash and personal data can leak. You cannot patch their server. You can make sure a leak there does not unlock anything else.
2. Credential-stealing malware. The LiteLLM payload shows the target list: saved browser passwords, session tokens, cloud keys, API keys. Anything stored in plain form on a machine that ran the bad package was fair game.
3. Better phishing and faster attacks. When exploitation windows shrink from weeks to hours, the old advice of "update when convenient" stops being safe, and fake login pages are cheaper to produce at volume.

There is a fourth route that is still thin on facts. Business Insider, citing Axios, reported that US banks were finding AI agents inside customer accounts, some apparently acting without their owners' instructions. We have not confirmed that from a primary source, but it is a good reason to check which apps and AI tools you have given access to your accounts.

The defences below are chosen to break each of those routes.

The priority list: do these first

Work through this in order. The first four items cover most of the risk, and none of them requires a technical background.

1. Secure your primary email account first. Whoever controls your email can reset almost every other password. Give it a unique long password, turn on the strongest second factor it supports, and review recovery options.
2. Use a password manager and give every account a unique password. A breach at one site then costs you one account, not twelve. Start with email, banking, your domain registrar, cloud and hosting accounts, social accounts you rely on for work, and any crypto exchange.
3. Move your important accounts to passkeys or a hardware security key. Details in the next section.
4. Update your browser, operating system and phone now, and turn on automatic updates. Browser vendors shipped certificate revocations within days. Updates are how you receive that protection.
5. Remove weak recovery paths. Delete old recovery emails you no longer control, replace security-question answers with random strings stored in your password manager, and prefer an authenticator app or key over SMS wherever you can.
6. Save recovery codes offline. Print them or store them somewhere that is not the device you would lose. Lockout is the real cost of strong security, so plan for it.
7. Check what is logged in. Review active sessions and connected apps on your email, Google, Apple, Microsoft, GitHub and social accounts. Revoke anything you do not recognise.
8. Turn on login alerts for email, banking and your registrar.
9. Audit connected apps and AI tools. Remove integrations you no longer use, and do not give an AI agent broad access to email, banking or cloud accounts unless you can limit and revoke it.
10. Put smart-home devices on their own network. Most routers offer a guest network. Cheap connected devices are rarely patched, so keep them away from your laptop and phone. Figma's John Milinovich gave this advice alongside two-factor authentication on everything, a password manager and an identity-protection service.

Passkeys and hardware keys: why they matter now

Most account takeovers still begin with a stolen or phished password. A one-time code from SMS or an app helps, but a fake site can ask you for that code and pass it along in real time.

Passkeys and FIDO2 hardware keys (such as YubiKeys) are tied to the real website address. If you land on a look-alike domain, the key simply does not answer. That removes the whole class of phishing in which you are tricked into typing a secret. Security-minded users on social media were describing the same benefit this week: one person who set up keys on a dozen critical systems said it let them ignore a whole genre of phishing emails.

How to do it well:

• Buy two keys and register both on every service that allows it. Keep one on you and one in a safe place. Losing your only key is the most common way people lock themselves out.
• Set a PIN on each key. It stops a thief who finds a key from using it.
• Start with email, your password manager, your registrar and your cloud accounts. Add others over time.
• If a service offers passkeys, use them. Phones and password managers can store passkeys, which is more convenient than a physical key and still resistant to phishing. A hardware key is the stronger option for your most important accounts.
• Know the limits. A key does not help if malware on your device steals a session that is already logged in, and it does not protect you if the password manager itself is compromised. It is one layer, not a shield.

One more note: researchers discussed a vulnerability affecting attestation in some YubiKey devices this week. Posts from people who looked at it described the practical security benefit of the keys as largely unchanged. Read Yubico's advisories for your model and update firmware where offered, rather than relying on a social-media summary.

Rotate credentials in the right order

If you have not changed important passwords in a while, or reused any of them, rotate them. The order matters:

1. Email accounts
2. Password manager master password
3. Domain registrar, DNS and hosting
4. Cloud and developer accounts (GitHub, GitLab, AWS, Google Cloud, Azure)
5. Financial and crypto accounts
6. Everything else

Change passwords from a device you trust, not the one you suspect. Sign out of all other sessions after the change.

If you own a website or domain

Black October was, at its core, a server-side event. If you run a site on PHP, you are the person who can act fastest.

• Scan and patch continuously. Krebs told CNBC that smaller companies, which lack access to frontier AI models, should scan for vulnerabilities all the time, patch as fast as possible, or take legacy systems offline. Retire anything unsupported that you cannot patch.
• Update PHP to a patched release: 8.2.35, 8.3.36, 8.4.27 or 8.5.12 (or newer). On shared hosting, check your control panel for the PHP version and ask the host if you cannot change it yourself.
• Update your CMS, themes and plugins. WordPress, Drupal and Joomla were all named in the research. Remove plugins you do not use.
• Restrict admin access. Require a passkey or key on the admin login, limit login attempts, and where possible allow admin access only from known IP addresses.
• Lock your domain. Turn on registrar lock, use a registrar account protected by a hardware key, and keep contact emails current. A domain hijack is worse than a hacked page.
• Protect DNS. Limit who can edit records and enable change notifications.
• Keep backups that are off the server and test a restore once. A backup stored on the same compromised machine is not a backup.
• Look at your logs. Unexpected new files, unfamiliar admin users and odd outbound traffic are the signals. If you see them, restore from a known-good backup and rotate every secret the server could read.
• Put a web application firewall in front of the site if your host or CDN offers one. It does not replace patching, but it can block known exploit patterns while you update.

If you build software

The LiteLLM incident is a template for what to expect: attackers got a maintainer's credentials, abused the release process, and published a poisoned version that harvested secrets from the machines that installed it.

• Pin dependencies and use lockfiles with hashes. Do not let production builds pull "latest" automatically.
• Wait a short while before adopting a brand-new release of a critical dependency. The malicious LiteLLM packages were live for under two hours.
• Keep secrets out of developer laptops and CI where you can. Use short-lived tokens, scope them narrowly, and store cloud credentials in a secrets manager.
• Set spending limits and alerts on AI and cloud API keys. A stolen key is often used for compute within minutes.
• Protect your own publishing accounts (PyPI, npm, Docker Hub, GitHub) with hardware keys, and use trusted publishing where it is available.
• Separate browsers or profiles for development and personal browsing, so a compromised dev environment does not hold your personal saved passwords.

If you think you were exposed

Run through this quickly:

1. Did you install LiteLLM 1.107.0 or pull the affected container images on 21 October 2026? LiteLLM's guidance is to rebuild the environment and rotate every credential that environment could read. Do not just upgrade in place.
2. Treat anything the machine could access as exposed: browser-saved passwords, cloud keys, SSH keys, API keys and tokens. Rotate them, starting with the highest-value ones.
3. Sign out everywhere and revoke connected apps and tokens.
4. Check billing and access logs for cloud and AI accounts for unfamiliar usage.
5. Watch for follow-on phishing. Leaked data is often used for convincing, personalised messages weeks later.
6. Tell people who rely on you (your team, your customers) if their data may have been involved.

Make yourself a harder target to research

Before attackers phish or impersonate someone, they research them. The same open-source techniques covered across this site are what they use: your public email addresses, usernames, old accounts, data-broker listings, domain registration details and leaked-credential dumps. Reducing that footprint makes targeted attacks harder.

• Use a different email alias for important accounts than for newsletters and social profiles.
• Turn on WHOIS privacy for domains you own.
• Search for yourself the way an investigator would, then remove what you do not need public. Our guide on how to hide yourself from OSINT walks through it.
• Be wary of any message that creates urgency around a login, invoice or security alert. Go to the site directly instead of using the link.

Keep it in proportion

This week produced a lot of alarm. Business Insider noted viral posts on X and Reddit predicting a global AI swarm attack, complete with stockpiling checklists. That is not what the evidence shows. The incidents are serious, but the advice from the people closest to them is about basics: patch quickly, use strong sign-in, limit access, and keep backups. Anthropic's Jason Clinton framed it as asking for help "not for AI, for the basics", and said defenders need to hold on for roughly 18 months.

A habit from OSINT applies here. Viral summaries of this story also listed further incidents, including app bugs and bank breaches. We could not confirm those from primary sources, so they are not in this article. Before you act on or share a claim, find the vendor advisory, the regulator notice or the original researcher's write-up.

What we do not know yet

This story is moving. Estimates of how many sites were exposed vary. Reports differ on how widely the techniques have been used in real attacks: Cloudflare saw large numbers of exploitation attempts, while other coverage says there is no evidence yet of attacks built with the same AI techniques. Details of some incidents may change as investigations finish. Check the vendor advisories linked below for current guidance, and treat any single tally of affected sites as approximate.

Frequently asked questions

Does Black October mean my personal accounts were hacked? Not by itself. The incidents were flaws in server software, a poisoned software package and a weak certificate authority. Your exposure depends on whether services you use ran vulnerable software, or whether you installed the affected package. The safe response is the same either way: unique passwords, phishing-resistant sign-in and prompt updates.

Should I stockpile supplies or go offline? No. Nothing in the reporting supports that. The practical response is the unglamorous one: update your devices, use a password manager, turn on phishing-resistant sign-in and keep backups.

Is a YubiKey enough to protect me? It is one of the strongest single steps against phishing and password theft, but not a complete defence. It will not stop malware that steals an active session, and it does not protect an account whose recovery options are weak. Pair it with a password manager, updates and good recovery settings, and register two keys.

Should I use passkeys or a hardware key? Both resist phishing. Passkeys are more convenient and are stored on your devices or in your password manager. A hardware key is a good choice for your most critical accounts, such as email, your password manager and your registrar. Many people use passkeys broadly and keys for the critical few.

Do I need to change all my passwords? Change the high-value ones first (email, password manager, registrar, cloud, financial) and any password you have reused. If a password is unique and strong and the service has not reported a breach, it can wait. If you installed the affected LiteLLM version, rotate everything that environment could access.

I run a WordPress site. What should I do today? Check your PHP version and update to 8.2.35, 8.3.36, 8.4.27 or 8.5.12 or later. Then update WordPress core, themes and plugins, remove unused plugins, protect the admin login with a key or passkey, and make sure you have a recent off-server backup.

Are the certificate problems something I need to fix? For most users, no. Chrome and Firefox revoked the affected eMudhra keys through their normal update channels, so keeping your browser updated is the action. If you manage systems that use their own trust stores, such as Java or Python environments, watch for vendor updates and install them.

Why are people saying patching is not enough anymore? Because AI shortens the time between a flaw being found and being exploited. Several executives quoted in the press argued that defenders cannot patch everything quickly enough. The practical answer is layers: automatic updates, phishing-resistant sign-in, least-privilege access, off-site backups and credentials that expire quickly.

Sources

• The Guardian, "'Black October': cybersecurity experts sound the alarm over new era of AI-enabled hacking" (23 October 2026)
• CNBC, AI's 'Black October': Critical PHP flaws set off global security scramble (23 October 2026)
• Business Insider, "Should we panic and turn off the internet?" (Alistair Barr, 23 October 2026)
• Wikipedia, Black October (information security)
• Wiz, PHP2Shell: AI uncovers critical RCE vulnerabilities
• Wiz, LiteLLM compromised, steals credentials
• LiteLLM, Security incident, 21 October 2026
• Google Security Blog, Protecting Chrome and Android users from compromised eMudhra intermediate CA certificates
• Mozilla Security Blog, Protecting Firefox users following the compromise of eMudhra intermediate certificates
• The Hacker News, Critical GitLab flaw (CVSS 10.0)

This article is educational and general. It is not a substitute for advice tailored to your organisation. Follow the official advisories from your software vendors.