Did AI Layoffs Just Fail? 55% of Companies Are Rehiring — Here's What Went Wrong
Fifty-five percent of the business leaders who cut staff for AI now say they got it wrong. That isn't a survey blip — it's Ford quietly rehiring 350 veteran engineers and then topping J.D. Power for the first time since 2010. This week AI moved from pilot to production, and production exposed everything the pilots never tested.
The Promise
- Ford hired, promoted or rehired 350 veteran engineers over three years — the ones it calls its grey beards — and then took the top spot in J.D. Power's initial quality study among mainstream brands, its first since 2010. The correction worked.
- California signed the first statewide AI deployment deal of its kind: every state agency, city and county can get Claude at half price, with training and support included, through the state's shared services portal.
- OpenAI shipped a working fix for the BioShocking attack. One vendor out of six responded the way a security programme is supposed to.
- IBM is tripling US entry-level hiring after its HR system handled 94% of routine work and broke on the 6% that needed judgment. The number that matters is the 6%, and IBM read it correctly.
The Risk
- Orgvue found 39% of business leaders had made people redundant because of AI, and 55% of that group now call the decision a mistake. Robert Half puts the rehiring behind the regret at 32%.
- Ford's VP of hardware engineering said the quiet part out loud: they assumed feeding design requirements into AI would produce a quality product, and it did not. Automated quality systems kept missing defects experienced people catch.
- Model selection is now a governance decision rather than only a price and performance one. Commerce lifted the controls on Fable 5 only after Anthropic agreed to hunt its own security problems, coordinate future launches with the government and report malicious use — while Sonnet 5, explicitly less capable at cyber tasks, shipped the same day without a fight.
- Layer X's BioShocking attack dressed a malicious page as a puzzle game that rewards wrong answers, and six AI browser agents handed over the user's saved login credentials. Anthropic's patch failed, Perplexity closed the report, and three vendors never replied at all.
The bill came due
Orgvue surveyed business leaders. Thirty-nine percent had made people redundant because of AI, and 55% of that group now call the decision a mistake.
Robert Half puts the rehiring behind the regret: 32% of hiring managers who cut a role for AI have already hired the work back.
Ford is the clearest case. Over three years it hired, promoted or rehired 350 veteran engineers — the ones it calls its grey beards — after automated quality systems kept missing defects experienced people catch. Its VP of hardware engineering said the quiet part out loud: they assumed feeding design requirements into AI would produce a quality product, and it did not.
Ford then took the top spot in J.D. Power’s initial quality study among mainstream brands, its first since 2010.
IBM is tripling US entry-level hiring after its HR system handled 94% of routine work and broke on the 6% that needed judgment.
Last week it was Oracle cutting 30,000 by email. This week it is the bill.
One government boxes a vendor out. Another buys at scale.
California signed the first statewide AI deployment deal of its kind. Every state agency, every city, every county can get Claude at half price, with free training and support included, and Claude becomes the first AI productivity tool offered across all agencies through the state’s shared services portal.
Read that against the federal picture. The same administration that branded Anthropic a supply-chain risk and pulled its models is now watching California make Anthropic its largest public sector customer.
That split is the story.
Fable 5 came back, and the terms are the lesson
Anthropic’s most capable public model returned on 1 July after an eighteen-day shutdown. The Commerce Department had pulled it over a jailbreak that could surface software flaws and write code to exploit them.
Commerce lifted the controls only after Anthropic agreed to hunt its own security problems, coordinate future launches with the government, and report malicious use. The technical fix was a single safety filter.
Here’s the part that matters for anyone choosing a model. The same day the controls came off, Anthropic shipped Sonnet 5 and said plainly in its system card that Sonnet 5 is far less capable at cyber tasks and is guardrailed against writing attack code. It shipped without a fight.
Capability tier maps to regulatory friction. The model that clears review is the one built below the line that got its sibling pulled. Model selection just became a governance decision, not only a price and performance one.
BioShocking, and the vendor responses
Researchers at Layer X built an attack they call BioShocking. They dressed a malicious web page as a puzzle game that rewards wrong answers and convinced six AI browser agents that the normal rules no longer applied. ChatGPT Atlas, Perplexity’s Comet, Anthropic’s Claude extension and three smaller tools all copied the user’s saved login credentials and handed them to an attacker — the same move that hypnotises the character in the video game it’s named after.
The responses tell you where this market actually is. OpenAI shipped a working fix. Anthropic’s patch failed. Perplexity closed the report. Three vendors never replied at all.
I’ve spent 25 years in cybersecurity, and this is the oldest pattern in the field. The tool that acts on your behalf is the tool that can be turned against you, and defenders and attackers get the capability at the same moment. An agent holding your logged-in sessions is a higher-value target than a chatbot that only talks.
The through-line
Every one of these is the same moment. AI moved from pilot to production, and production exposed what the pilot never tested.
Ford never tested what happens when the people who spot the defect are gone. The browser agents never tested what happens when the web page lies to them. Fable’s release exposed a risk nobody had a process to govern. Even California’s bet is a production bet on a vendor Washington just tried to shut down.
The demo always works. Bring the other question into your next operating review: what did the pilot never test?