Updated October 2026. This post originally covered 2024; I’ve rewritten it for where security stands now.
Every year someone says cybersecurity is “changing faster than ever”, and every year it’s true. What feels different in 2026 is where the change is coming from. The biggest shifts aren’t new malware families. They’re AI systems that can take actions, attackers going after identity instead of infrastructure, and a cryptography deadline that is finally close enough to plan for.
Here are the trends I’m watching as a security researcher, and what each one means in practice.
1. Agentic AI turns prompt injection into real-world actions
A chatbot that gives a wrong answer is embarrassing. An AI agent that can browse the web, read your email, call APIs and run code is a different risk. If an attacker can slip instructions into a web page or a document the agent reads, that “prompt injection” can now trigger actions, such as sending data out, changing records or running commands.
This is no longer theoretical. HiddenLayer’s 2026 AI Threat Landscape Report found that around one in eight reported AI breaches was linked to agentic systems, and 76% of the security leaders surveyed see “shadow AI” (tools adopted without security review) as a definite or probable problem, up from 61% a year earlier.
What to do: treat every agent like a new employee with too much access. Give it the least privilege it needs, require human approval for sensitive actions, and assume anything it reads from outside can contain hostile instructions.
2. The AI supply chain is the new open-source supply chain
Teams pull models, datasets and AI libraries from public repositories the same way they pull npm or PyPI packages, and attackers have noticed. In the same HiddenLayer survey, malware hidden in public model and code repositories was the most-cited source of AI-related breaches (35%), while 93% of organisations still rely on those repositories.
What to do: scan models and packages before use, prefer safe model formats over ones that can execute code when loaded, and pin versions the same way you would for any dependency.
3. AI on both sides of the fight
Attackers use AI to write convincing phishing in any language, clone voices for fraud calls and speed up reconnaissance. Defenders use it to triage alerts, summarise incidents and spot anomalies across huge volumes of logs. Neither side has a lasting advantage. The organisations doing well are the ones that use AI to remove repetitive work from analysts, not to replace their judgement.
For individuals, the practical lesson is simple: a familiar voice or a polished message is no longer proof of anything. Verify urgent requests through a second channel. I’ve written about one real example in this WhatsApp police scam.
4. Identity is the real perimeter, and passkeys are winning
Most breaches I see discussed today start with a stolen session, a phished password or an over-permissioned account, not with an exotic exploit. That’s why passwordless sign-in matters. The FIDO Alliance estimated in May 2026 that around 5 billion passkeys are in use, and 75% of people surveyed have enabled one on at least one account.
The gap is in the workplace: 57% of organisations still rely on phishable methods for employees’ everyday sign-in. Closing that gap is one of the highest-impact things a security team can do this year.
What to do: turn on passkeys wherever your accounts support them, starting with email and banking. My guide to protecting your personal information online covers the other basics.
5. Post-quantum cryptography moves from research to project plan
Quantum computers that can break today’s public-key encryption don’t exist yet, but the migration deadlines do. NIST’s guidance (IR 8547) calls for RSA-2048 and 256-bit elliptic-curve cryptography to be deprecated by 2030 and disallowed after 2035, and the US CNSA 2.0 rules require new national security system purchases to support quantum-resistant algorithms from January 2027.
The reason to care now is “harvest now, decrypt later”: data encrypted today can be stored and decrypted years from now. Anything that must stay confidential for a decade is already at risk.
What to do: start with an inventory. You can’t migrate cryptography you don’t know you’re using.
6. Zero trust becomes the default, not a project
“Never trust, always verify” was a buzzword a few years ago. Today it’s simply how modern access works: every request is checked against identity, device health and context, whether it comes from inside the office network or not. The trend now is less about buying a zero-trust product and more about applying the principle everywhere, including to AI agents and machine identities.
7. Regulation and disclosure keep tightening
Privacy and breach-reporting rules continue to expand worldwide, including India’s Digital Personal Data Protection Act. There’s still a gap between intention and practice: in HiddenLayer’s survey, 85% of leaders supported mandatory breach disclosure, yet 53% admitted to holding back a breach report for fear of backlash. Expect regulators to keep narrowing that gap.
What this means if you work in security
If you’re building a career or hunting bugs, the skills worth investing in are shifting:
- Testing AI applications: prompt injection, insecure tool use and data leakage in LLM features are now common findings in real programs.
- Identity and access flaws: broken authorisation, session handling and OAuth mistakes remain some of the most valuable bugs.
- Cloud and supply-chain thinking: understanding how code, packages and models get from a repository into production.
The fundamentals haven’t gone anywhere. Curiosity, careful testing and clear reporting still matter more than any single tool. If you’re just starting out, my post on building a career in cybersecurity is a good next read.
