OPENAI IS NOW IN THE BUSINESS OF SELLING HACKING TOOLS AND CALLING IT DEFENSE
OpenAI just updated GPT-5.5-Cyber, its specialized cybersecurity model that sits behind a velvet rope and is only available to vetted partners and government agencies. The new version scored 85.6 percent on CyberGym, a benchmark that measures whether an AI agent can reproduce known software vulnerabilities in controlled testing environments. The previous version scored 81.8 percent. So yes, OpenAI’s AI is now measurably better at hacking things than it was three months ago.
But here is the part that actually matters. OpenAI is now opening this up to outside companies through something called the Daybreak Cyber Partner Program. Vetted cybersecurity vendors can plug GPT-5.5-Cyber directly into the products they sell to their own customers. Which means the circle of organizations with access to a state-of-the-art vulnerability-finding AI just got significantly larger.
OpenAI is very careful with the language here. Everything is “authorized cybersecurity work” and “trusted partners” and “defenders.” That framing is doing a lot of heavy lifting. What the model actually does, stripped of the marketing, is analyze large codebases, identify security-relevant components, validate likely vulnerabilities, and develop patches to fix them. That is an extraordinarily powerful capability. It is also, depending on who is running it, a very capable tool for finding and exploiting holes in other people’s software.
To give credit where it is due, OpenAI is also funding something called Patch the Planet alongside Trail of Bits and HackerOne, which is genuinely trying to help the open source community find and fix vulnerabilities before bad actors do. That is a real program doing real work. The optimistic version of this story is that defenders now have a tool powerful enough to outrun attackers, and the internet gets meaningfully safer as a result.
The less optimistic version is this: OpenAI has now signed cyber partnership agreements with Australia, Canada, France, Germany, Japan, Poland, South Korea and EU institutions. Every one of those governments now has access to a model that can score 85.6 percent on a benchmark for reproducing known vulnerabilities. Every security vendor in the Daybreak program is another organization holding keys to the same capability. The attack surface for this technology is expanding in real time, and the pace at which the rules governing it can be written and enforced is considerably slower than the pace at which the technology is being deployed.
The assumption baked into all of this is that everyone in the program stays vetted, no one gets breached, and no partner ever decides that a lucrative contract with an ambiguous client is worth bending the definition of “authorized.” History does not strongly support that assumption. But the technology exists now, so the question is not whether it gets deployed but how carefully. OpenAI is at least asking the question out loud, which is more than can be said for some of its competitors.
Source: Axios: OpenAI gives GPT-5.5-Cyber more powerful cybersecurity capabilities
GOOGLE BUILT AN AI SEARCH ENGINE AND PEOPLE ARE RUNNING BACK TO THE INTERNET OF 2005
Something that was not supposed to happen is happening. DuckDuckGo, the privacy-focused search engine that most people treat as a curiosity rather than a real alternative, is seeing install rates jump by up to 40 percent per week. Not 40 percent year-over-year. Per week. For a search engine that has been unable to crack 2 percent of the US market for the entirety of its existence, that is a stunning number.
This is not a story about DuckDuckGo suddenly getting good. DuckDuckGo has not invented a better algorithm. What is happening is much simpler and more interesting. Google held its I/O developer conference and announced the biggest overhaul of Search in twenty-five years. Instead of a list of links to websites, you now get an AI-generated summary that pulls from sources you cannot easily see or verify, and there is no obvious way to opt out of this experience. For a large and apparently growing segment of Google’s user base, this is a worse product than what they had before.
CNBC’s piece on Google’s cracking dominance is worth reading carefully. Google CEO Sundar Pichai said this transformation is “the most significant change in Search since the introduction of the link.” He is probably right about that. The question is whether the change is an improvement. Google’s answer is yes, obviously, look at the engagement metrics. The answer from people downloading DuckDuckGo at 40 percent weekly growth is somewhat different.
DuckDuckGo CEO Gabriel Weinberg put it plainly: Google is “force-feeding AI with no way to opt out.” The company launched a specific no-AI search browser extension and noai.duckduckgo.com for users who want search results that are just, you know, results. Links to websites. The way it used to work. The fact that there is meaningful demand for “search but without the AI” in 2026 is a data point that should make Google’s product team uncomfortable.
Now for the honest context. Google still controls 90 percent of the US search market. Its stock has more than doubled in the past year. Revenue growth in Q1 2026 was the fastest since 2022. By every financial measure, Google is winning. The 40 percent weekly DuckDuckGo growth is happening off a very small base and DuckDuckGo is still at roughly 2 percent market share.
But the gap between “winning financially” and “users actively prefer the product” is now visible, and these things have a way of accelerating. Google’s entire market position was built on being the best tool for finding things on the internet. It was not built on being the most profitable company or having the best AI. If users decide in sufficient numbers that Google’s AI-first search is actually worse for their specific purposes, the network effects that have protected Google for twenty years start working in the opposite direction. Not quickly. But the first 40 percent weekly growth crack in a twenty-five-year wall is worth paying attention to.
Source: CNBC: Google’s online dominance is showing signs of cracking in AI era
ORACLE JUST FIRED 21,000 PEOPLE AND THE REASON IT GAVE REGULATORS WAS THE AI IT SELLS
Oracle reduced its workforce by about 13 percent over the past year and now employs 141,000 full-time workers, down from 162,000. In its annual regulatory filing, Oracle cited AI deployment as the reason for the cuts. Read that again. A company that sells AI cloud infrastructure to other companies used that same AI to eliminate roughly one in eight of its own employees, and then wrote that explanation down in documents submitted to the Securities and Exchange Commission.
This is not the shocking part. The shocking part is that it is now normal. According to CNBC, employers cited AI as the primary reason for almost 40 percent of announced job cuts in May 2026. In January of this year, AI was the stated reason for 7 percent of cuts. By February it was 10 percent. March, 25 percent. April, 26 percent. May, 40 percent. That is not a trend, that is an acceleration. The total number of jobs officially attributed to AI in just the first five months of 2026 is 87,714. The entire 2025 total was 54,836.
The companies doing the cutting are not struggling businesses looking for ways to survive. Amazon cut 16,000 corporate jobs following 14,000 cuts in late 2025. IBM has eliminated more than 15,000 US positions since September 2024. Intuit is cutting 3,000 people, about 17 percent of its workforce, in a “restructuring centered on AI.” Snap cut a thousand employees and 300 open positions. The CEO of Wix announced a 20 percent workforce reduction. These are profitable, growing companies. They are not cutting because they have to. They are cutting because the return on labor has shifted and the math now works differently with AI doing portions of what those employees were doing.
TechCrunch published a piece this week calling the AI layoff wave “a powder keg” and the framing seems accurate. The standard response from tech executives is that AI creates new jobs as it eliminates old ones. That may be true over a long horizon. A 2026 Motion Recruitment study found that AI adoption is specifically slowing hiring for “generalized IT roles” while demand for AI-specific positions is strong. The person who got laid off from Oracle is not automatically qualified for the AI engineering role that replaced their function. That gap is real and it takes years to close, not months.
Eighty-seven thousand jobs in five months, at an accelerating pace, disproportionately concentrated in accessible entry-level and mid-level positions, with the gains going to people who were already highly technical. That is a political story as much as an economic one. The companies racing to automate are not thinking about the political consequences of the pace. Someone else will think about it for them.
Sources: CNBC: Oracle sheds 21,000 roles | TechCrunch: The running list of major tech layoffs citing AI
THE STARTUP THAT BUILT A TOOL TO DETECT AI CONTENT JUST SOLD ITSELF TO AN AI COMPANY. NOBODY SEEMS TO FIND THIS WEIRD.
GPTZero is a three-year-old company that Princeton student Edward Tian built as his senior thesis and then turned into a real business. The product detects AI-generated writing. Teachers use it to catch students submitting ChatGPT essays. Editors use it to screen submitted work. Employers use it to flag AI-written job applications. It amassed 19 million registered users and grew to $30 million in annual recurring revenue. On a $13.5 million total raise, that is an impressive outcome.
GPTZero was acquired this week by Superhuman. Superhuman is the company that used to be called Grammarly until it bought an AI email product also called Superhuman and rebranded, which is a naming situation that remains confusing no matter how many times you explain it. The point is that Superhuman, the company, is an AI writing tool. It uses AI to help people write better. It also, as it turns out, already had its own AI detection tool built into its platform.
So to summarize: an AI writing company that helps users produce AI-assisted content just bought the most popular tool for detecting AI-assisted content. And when asked to explain this, the company’s official statement was that “two AI detectors are better than one.”
That statement is technically true in the same way that a bank saying “two vaults are better than one” is technically true. It is not wrong. It also does not fully address the interesting structural question, which is what it means for GPTZero’s independence and mission when it is owned by a company whose core product moves in the opposite direction.
Grammarly’s AI detection tool, per the TechCrunch writeup, was “designed to help users, often students, determine whether their writing appears AI-generated, then revise it so it doesn’t.” That is not the same as detecting AI writing to prevent it. That is detecting AI writing to help it pass. GPTZero’s mission was to help humans identify when they are being fooled. Grammarly’s existing tool was helping users fool the detectors. Those two things are now inside the same company.
Edward Tian is apparently staying involved, and GPTZero will continue running as a product. Maybe both tools remain independent and the combined company genuinely is better at detection than either was alone. Or maybe the dynamics of being owned by an AI writing tool company subtly shift the priorities over time. Corporate acquisitions have a way of aligning incentives whether the founders intended it or not.
The broader irony is that this acquisition is a perfect little model of the AI economy in 2026. The referees are getting bought by the teams they were hired to watch. The auditors are merging with the firms they were auditing. The companies that built the problems and the companies that built the solutions are consolidating into the same entities. This is not a conspiracy. It is just what markets do when a new technology moves faster than the institutions designed to manage it.
Source: TechCrunch: Superhuman acquires AI detection startup GPTZero