You open an email. The grammar is clean. The tone matches your company's communication style. It uses your first name, references a real project you're working on, and the sender appears to be your manager. Your gut says nothing. You click.

That's not a hypothetical. That's what phishing looks like in 2025 — and it's why the advice you've been given for the last decade is now dangerously outdated.

The Shift

For years, security trainers told people: "Look for spelling mistakes. Grammatical errors. Awkward phrasing." That wasn't bad advice — it worked. Bad actors operating at scale made errors because phishing was often handled by people working in their second or third language, building campaigns quickly, with no editorial eye.

But here's the part nobody talks about enough: some sophisticated attackers kept the typos on purpose. A poorly written email self-selects for highly susceptible victims. If you click despite obvious errors, you're more likely to follow through on the rest of the scam. The sloppiness was a feature.

Now, that entire calculation has been erased. LLMs write clean, fluent, contextually appropriate prose — at zero marginal cost, in any language, in any tone, at any volume. The last reliable surface-level signal is gone.

4,000% Rise in malicious emails since ChatGPT launched

60% Higher click rate on AI-generated vs human-written phishing

<30s Time to generate a convincing, personalized lure with AI

Core Insights

01 — THE PERSONALIZATION PROBLEM

It's not just better grammar. It knows who you are.

LLMs don't just fix sentences — they enable context-aware targeting at scale. A threat actor can feed your LinkedIn profile, your GitHub commits, your public Slack threads, and your company blog into a model and generate an email that reads like it came from a colleague. It references real projects. It matches real timelines. It mirrors your manager's communication style. This level of personalization used to require a human intelligence operative and days of work. Now it takes a prompt and ten minutes.

02 — THE DEMOCRATIZATION EFFECT

Nation-state quality attacks are now available to amateurs.

Spear phishing — highly targeted, researched, personalized attacks — was once the domain of sophisticated threat actors. APT groups. State-sponsored hackers. People with time, resources, and tradecraft. That barrier is gone. A moderately motivated attacker with a free API key and basic scripting knowledge can run campaigns that, six years ago, would have required an entire red team. The skill floor collapsed. The threat ceiling didn't change. That gap is the problem.

03 — THE CHANNEL EXPANSION

Email was just the beginning. Language is everywhere now.

The conversation around AI-powered phishing stays focused on email. That's the wrong scope. AI-generated social engineering is now showing up in Slack DMs, LinkedIn messages, Teams chats, SMS, GitHub issues, and even voice calls using cloned audio. Wherever language is used to build trust, the attack surface exists. The people most at risk aren't necessarily those who check email carelessly — they're the ones who never considered that a Slack message from "IT" might not be from IT.

04 — THE VERIFICATION VACUUM

We built trust infrastructure for a pre-AI world.

Most security awareness training still teaches people to evaluate content — check the sender address, look for urgency, inspect links. That model assumed an imperfect attacker. It assumes that if you look closely enough, something will be off. That assumption no longer holds. The content can now be flawless. What matters isn't reading the email carefully — it's verifying the request through a separate, trusted channel. The attack is a language problem. The defense has to be a process problem.

The uncomfortable truth: Telling people to "be more careful" with email is now roughly as useful as telling them to "trust their gut." You can't out-careful a model that was trained on billions of human interactions and can generate perfect prose in milliseconds. Awareness is still necessary. It's just no longer sufficient.

Real-World Relevance

In early 2024, security researchers at IBM demonstrated AI-generated phishing campaigns that outperformed human-crafted ones in controlled tests — higher open rates, higher click rates, more convincing urgency framing. No prompting expertise required. Just a model, a target profile scraped from public sources, and a template.

Deepfake audio is now being used in vishing attacks — phone calls where the voice sounds like a CFO, a legal contact, or an IT administrator. A finance team in Hong Kong transferred $25 million after a video call that appeared to include multiple real executives. Every person on that call was AI-generated.

These aren't edge cases anymore. They're the leading edge of what becomes standard in 18 months.

What To Do

Content quality is no longer a reliable trust signal. You need to build the habit of separating what someone is saying from whether the request is legitimate. Those are different questions. Here's where to start:

Verify through a separate channel — always. If you receive an unexpected request involving credentials, money, access, or urgency, verify it via a method that wasn't used to send the request. Call the person directly. Use an established number. Not the one in the email.

Stop evaluating tone, start evaluating context. Ask: Was I expecting this request? Does the timing make sense? Would this person normally send this through this channel? Unusual context is the new typo.

Apply the same skepticism to every channel. A Slack message from IT asking for your credentials deserves the same scrutiny as a suspicious email. Language-based trust now needs to be earned through process, not prose.

For organizations: implement process-level controls. Require out-of-band verification for any financial transaction, credential reset, or access grant. No email — however perfect — should be sufficient authorization on its own.

Use hardware-based MFA where possible. AI-powered phishing can capture session tokens through adversary-in-the-middle proxies. Phishing-resistant MFA (FIDO2, passkeys, hardware keys) doesn't rely on the user catching anything — the protocol does the verification.

The phishing email you can't spot isn't a failure of attention. It's a failure of the model we built trust on. The old game was: find the flaw. The new game is: verify the request — regardless of how flawless it looks.

The attacker's job got easier. Your process has to get better to compensate.

SKBDecoded  ·  skbdecoded.beehiiv.com  ·  Made for curious minds in cybersecurity