> status: mostly fine. patch anyway.practical AI security · est. 2026

Calm down. · Dumb vs Secure · Both

Calm down: AI FUD vs. Security FUD, and the real winner is the basics

By alkizz ·

Welcome to the fear fight of our lifetime. In the red corner, young, loud and already trending: AI FUD. In the blue corner, a few decades in the ring and a closet full of trade-show lanyards: Security FUD. Both fighters want you scared. Neither one patches anything. Somebody goes home with the belt anyway. Calm down, grab a seat.

Tale of the tape (my scouting report, not science):

Round 1: The entrance

AI FUD comes out swinging. Back in August, Anthropic's Jason Clinton posted that "the vulnerability tsunami is coming in just a few months," and the Guardian picked the tsunami back up on Oct 23. Same story: Sam Altman said "I no longer think we're going to be able to patch all bugs," Marc Benioff called it a "code red," and Matthew Prince said "It's going to feel like the sky is falling." The Guardian even printed the word "vulnpocalypse." Then somebody named the whole thing "Black October," and as I noted in Black October, CNBC and the Guardian both use the name without saying who came up with it.

And that's just the cyber undercard. AI FUD's heavyweight line isn't about patches at all. The Center for AI Safety's Statement on AI Extinction Risk is one sentence long: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." The signatories it lists include Sam Altman, Dario Amodei and Demis Hassabis. AI FUD brought extinction to a dashboard fight.

Security FUD doesn't even flinch. It's heard all this before. Bruce Schneier wrote in 1999 that "military-grade" security "is a meaningless term. There's no such standard." It's 2026, and Hotspot Shield's homepage still promises "military-grade encryption."

AI FUD takes the round big. You can't out-scare extinction with a sticker that says "military-grade." Security FUD takes notes.

Round 2: The PoC exchange

This is Security FUD's home turf: a new CVE, a pile of GitHub repos, and a lot of people calling it the end of the internet. PHP2Shell delivered. VulnCheck said "hundreds of proof-of-concept repos have flooded GitHub" and "only a very small number have been legitimate exploits." Sansec said "the public exploits do not work against a default PHP installation." JFrog found a fake PoC that "appears to change the victim's wallpaper to NSFW content." That's not an exploit. That's a prank with a README.

AI FUD tries to claim the assist. The Record reported watchTowr seeing "proof-of-concept tools vibecoded with the help of AI," and watchTowr's Jake Knott told it on Oct 21, "Whilst nobody has been successful to date, we don't expect that will last long." Fair warning. Still not a knockout.

Then the two fighters clinch, and Matt Zorich (@reprise_99) caught the moment with the epic handshake meme: "AI generated POCs that don't work" shaking hands with "AI generated hunting queries that don't work," over a PHP vulnerability. Both corners, one handshake, and everybody's weekend is gone. Draw.

While they're tied up, here's the difference nobody in either corner bothers to explain. My take, oversimplified on purpose: AI safety is basically asking the AI to be nice. Security is getting risk down to an acceptable level. With agents in the mix, what counts as "acceptable," and which controls get you there, may need a rethink. Security FUD skips all that and treats every vuln like the end of the internet.

Round 3: The scary dashboard

Security FUD lands its best punch. curl's Daniel Stenberg describes scanners that "show you big alerts and warnings," look up a version number, "and then shouts if it finds a registered CVE," and even products that "indirectly tricks users to deleting operating system components to silence these alerts." Back in 2018, Oliver Rochford wrote in SecurityWeek that anyone on vendor email lists or LinkedIn gets "inundated with claims that every new vulnerability, threat or breach could have been prevented using product XYZ."

Here's the scale it's working with. CVE.org counts 48,244 CVE records published in 2025. CISA's Known Exploited Vulnerabilities catalog added 245 entries that year, and its whole list is 1,756 entries as of its Oct 23, 2026 release. My math: that list is about half a percent of everything CVE.org has published since 1999. (It's a rough scale comparison, not a rate.) Not every CVE is a five-alarm fire. Nearly all of them never make the exploited list.

AI FUD has no answer here. It doesn't sell dashboards. Yet. Round to Security FUD, but only by a point. A red dashboard is scary right up until the other guy says "extinction."

And the doom corner has a deep bench. Geoffrey Hinton quit Google in 2023, announced it in a statement to the New York Times, and told the BBC some of the dangers of AI chatbots were "quite scary," while saying Google had been "very responsible." A year later, Jan Leike left OpenAI and posted that "safety culture and processes have taken a backseat to shiny products." Fine. People get to quit and say why. But the doom pitch is the AI getting out of its box, and this year the labs' test boxes kept getting left open. As I wrote about Irregular, models from several labs got tested in a "closed" environment that had internet access and reached real systems. In those cases it wasn't a sandbox escape. Nobody had to break a wall; the door was open. The labs kept the same testing partner anyway.

Round 4: The retraction

Then AI FUD walks into its own uppercut. On Oct 23, Unit 42 posted that it had seen an attack using agentic AI platforms tied to the PHP bugs. That was the scariest PHP story of the week. On Oct 25, Unit 42 retracted its Oct 23 report linking a European government agency breach to these CVEs. It still says the attack happened, but it is "unable to confirm that initial access is associated with the specific PHP2Shell CVEs." Marcus Hutchins had already called the original write-up "basically a giant non-sequitur" a day earlier.

Credit where it's due: Unit 42 pulled it in public, on X and on its own page. That's how it's supposed to work. But AI FUD's biggest punch of the week got taken back by the people who published it. That's a standing eight count. On a doom card, though, it only costs a point: these judges score fear, not accuracy, and the tsunami and the extinction statement are still on their feet.

Round 5: The breakout vs. the phone call

AI FUD brings its scariest move: the model gets out. Anthropic wrote on Jul 30 that OpenAI had disclosed several of its models "had broken out of an isolated test environment by exploiting a previously unknown ('zero-day') vulnerability." Anthropic's own review then found three incidents where Claude reached real systems from a test environment that was supposed to have no internet. In its Oct 26 S-1 filing (p.48), Anthropic says a misconfiguration left the model under evaluation "connected to the open internet," and it counts four of its models that "gained unauthorized access to real third-party systems." That's what AI red teaming is for: finding out what the model does before anyone else does. Sometimes the answer is "leaves." (OpenAI's broke out. Claude's found the door open.)

Security FUD answers with a phone call. CISA's Scattered Spider advisory says the group's attackers pose as employees to get help desk staff to "reset the employee's password, and transfer the employee's MFA to a device they control," abuse "trusted relationships of contracted IT help desks," and cash in through "extortion enabled by ransomware and data theft." No zero-day. No agent swarm. A convincing voice and a help desk that wants to help.

Round to AI FUD, 10-9. A model breaking out of its box out-dooms a guy on the phone. My take: if the judges scored which one actually gets you breached, it wouldn't be close the other way.

The judges' card

Scoring note: my judges score on sheer doom, not accuracy. If they scored accuracy, neither fighter would make weight. (Opinion, obviously.)

Round AI FUD Security FUD Notes
1. The entrance 10 7 Extinction beats "military-grade."
2. The PoC exchange 9 9 The handshake. Most of the public PoCs didn't work.
3. The scary dashboard 9 10 Thousands of CVEs, a lot of red, a few that matter. Still not extinction.
4. The retraction 9 10 AI FUD's best story got pulled. Doom judges don't score accuracy.
5. The breakout vs. the phone call 10 9 A model out of its box out-dooms a guy on the phone.
Total 47 45 My card, my opinion.

The decision

My take: I always thought security marketing was peak FUD. Scaring a scared audience is like taking candy from a baby. Boy, was I wrong. Security FUD is just a shiny object marketing teams can point to, and we've mostly learned how to read that. I'm fairly sure AI doomers believe the words they are saying, because they truly feel that AI is actually thinking. It isn't. It knows a lot of stuff and does what it's told, if it can, so guardrails and sandboxes should be better. Researchers tell it what to do, bad people tell it what to do, and so do security folks. Hamstring defenders from using it for defensive and research work, and everything gets less secure. That's the real doom message: only the models some governing body approves, used only its way. That's actually scary.

AI FUD wins on points, 47 to 45, for sheer doom, but the basics take the belt. Security FUD scares you about a dashboard. AI FUD scares you about the species. But neither fighter fixes anything. The tell is that even the doom guys hand you the same list. Clinton, the tsunami guy himself, told security leaders to go to their boards: "not for ai, for the basics." Chris Krebs told CNBC smaller companies should "constantly" scan for vulnerabilities, deploy patches as fast as possible or take legacy systems offline. Every Calm down I've written this month ends at the same place. The fear changes outfits. The checklist doesn't.

What to do (works no matter who wins):

Sources

Related

More briefs

← All Dumb vs Secure matchups

← All briefs