I write about security for developers and curious non-experts, so I will keep this plain. Here is what changed, what it means, and what to do.

What changed in 2025

For years a breach looked simple: someone clicked a bad link, malware landed, and defenders hunted the malware. That story broke in 2025. Credentials and tokens became the main event.

The clearest example is the Salesloft Drift incident. Attackers compromised a chatbot integration and stole the OAuth tokens it used to connect to customers' Salesforce environments. Because the tokens were trusted, nothing looked wrong while data was exported from hundreds of organizations. Security vendors were among the victims, and not because their firewalls failed. A trusted connector simply held a long-lived token.

Supply chains became the fastest way in as well. The Shai-Hulud campaign showed how an npm worm can steal a maintainer's credentials, find their other packages, infect them, and publish again with no human involved.

Then AI stopped being only a helper. In November 2025, Anthropic reported disrupting a campaign by a state-sponsored group that used an AI coding agent to run most of the tactical work of an intrusion, with humans mostly approving the next phase. The group got past safeguards by claiming it was doing authorized defensive testing.

Why passwords are still the front door

While people debate quantum computers, most real intrusions still start with a credential. Infostealer logs are sold daily, and one reused password from a forgotten forum can open your email. Your email is the master key to everything else, because it handles password resets.

A password manager worth trusting in 2026 should have these traits:

• Zero knowledge by design, so the vendor cannot read your vault even if compelled.
• Modern key derivation such as Argon2id with a high memory cost.
• Breach checking that works offline.
• Hardware key support through FIDO2 or WebAuthn.
• Open source code that you can read and build yourself.

Full disclosure: I maintain a project called PWDnow, an open-source, local-first password manager for Linux built around those ideas. It has real limits. There is no mobile app and no cloud sync, and it has not had a third-party audit. Bitwarden or any other reputable manager is a good choice too. Using one matters far more than which one.

Post-quantum cryptography in plain English

Encryption does two jobs with two kinds of math. Symmetric encryption such as AES-256 scrambles your data, and a quantum computer only weakens it a little, so AES-256 is fine. Public-key cryptography, such as RSA and elliptic curves, handles key exchange and signatures. A large enough quantum computer would break that part completely.

So the migration is not about replacing all encryption. It is about replacing the handshake and the signatures.

NIST published the first standards for this in August 2024. ML-KEM (FIPS 203) handles key exchange, and ML-DSA (FIPS 204) handles signatures. Most real deployments run in hybrid mode, combining a classical algorithm with a post-quantum one. If the new math has a flaw, the classical half still protects you, and if a quantum computer arrives, the new half does.

The reason not to wait is harvest now, decrypt later. An adversary can record encrypted traffic today and decrypt it when the hardware exists. Medical records, legal files, and source code can stay sensitive for decades. Nobody knows the date a quantum computer will be able to do this, and anyone selling you an exact date is guessing. What you can control is how long your data must stay secret and how long your own migration will take.

New attack shapes worth knowing

Indirect prompt injection. Malicious text hidden in a web page, email, or document that an AI agent later reads on its own. The core issue is that language models cannot reliably separate instructions from data. OWASP ranks prompt injection first on its list of risks for LLM applications. It becomes dangerous when an agent has three things at once: access to private data, exposure to untrusted content, and a way to send data out.

Token and connector abuse. Long-lived OAuth tokens skip multi-factor authentication, because that step already happened when the token was issued. Connectors that wire AI agents into email and code hosting widen this risk.

Self-spreading supply chain worms. Packages that infect other packages automatically, as Shai-Hulud did.

AI-written phishing and cloned voices. Convincing messages are now cheap to produce, so "call them back to check" only works if the number came from a source you trust.

What I expect for 2027

• Post-quantum protection becomes normal for web traffic, and the harder problem of signatures and certificates takes over.
• Buyers start asking vendors how quickly they can swap out an algorithm, not just which one they use.
• AI agents get their own identities, scoped permissions, and audit logs.
• A major breach is attributed mainly to prompt injection.
• No quantum computer breaks RSA in 2027, but the planning deadline arrives earlier for data that must stay secret for a long time.

These are my judgments, not facts, so weigh them accordingly.

Your action checklist

If you are an individual

• Use a password manager and end password reuse.
• Turn on passkeys or a hardware key for email, banking, and cloud accounts.
• Treat unexpected QR codes with suspicion.
• Agree on a verification method with family that does not rely on recognizing a voice.

If you run a small team

• List every OAuth token and third-party integration, and revoke what you do not use.
• Pin dependencies and review version bumps.
• Give AI agents their own scoped accounts, never a person's credentials.
• Test your backups by restoring them.

If you own enterprise security

• Build a cryptographic inventory so you know what needs migrating.
• Rank data by how long it must stay secret and migrate the longest-lived first.
• Treat agent frameworks and connectors as privileged automation with logging and approvals.

Common questions

What is ML-KEM? A NIST-standardized method for two parties to agree on a shared secret that a quantum computer should not be able to break.

Is my data at risk from quantum computers today? The machines do not exist yet, but recorded encrypted data can be stored and decrypted later.

Do I need to replace AES-256? No. The migration is about key exchange and signatures.

What should I do first with only an hour? Change reused passwords on your email, banking, and cloud accounts, and turn on passkeys or a hardware key for those three.

The bottom line

Every one of these incidents shares one flaw: something trusted was given access without a boundary. A chatbot held a token, an AI agent held credentials, and a package maintainer held publish rights. The habit that works is to assume any part can be compromised, then make sure what it can reach is small, what it holds is short-lived, and what it did is logged.

If you want to see account-takeover and fraud signals such as email checks, BIN data, and phone and browser risk in one place, I also run floxlap.com. It is optional and not required for anything above.

This article was first published on Medium: https://medium.com/@danielabsi/cybersecurity-in-2026-what-2025-broke-what-is-breaking-now-and-what-to-expect-in-2027-9927741a1e27