Something Happened This Week. I'm Not Sure What to Call It.
- Rich Washburn
- 2 days ago
- 6 min read
Something Happened This Week. I'm Not Sure What to Call It.
I have been in and around security and technology long enough to be suspicious of hype. The word "singularity" gets thrown around like a party favor by people who want to feel like they are living inside a historic moment. Most of the time, the moment is smaller than they think. This week is different.
I am not going to tell you the singularity has arrived in the Kurzweil sense. That is a specific technical claim about recursive self-improvement exceeding human control, and we are not there. Not definitively. But something has shifted in the slope of this curve that I cannot look at casually. And I think it is worth walking through it event by event, because the individual items are remarkable enough. The fact that they happened in a compressed window is something else entirely.
Here is the list that stopped the internet:
April 7 — Glasswing finds zero-days. A lot of them.
Anthropic launched Project Glasswing as a closed partnership with about fifty major technology, cybersecurity, and financial organizations. The model at the center of it — Claude Mythos Preview — was a frontier reasoning system trained specifically for offensive security work. Within weeks, it had identified more than ten thousand high- and critical-severity vulnerabilities across every major operating system, browser, and core infrastructure platform. It generated working exploits for 83 percent of what it found.
Ten thousand zero-days. With working proof-of-concept code for most of them.
Anthropic kept the model tightly gated for obvious reasons. They weren't being modest. The thing that will rebalance the attack-defense equation is the same thing that could hand an adversary a master key to most of the software the world runs on. Glasswing is being used by defenders first. But the capability exists, and it is not going back in the box.
April 14 — Erdős #1196 falls.
GPT-5.4 Pro solved an open number theory problem that had been sitting on the Erdős list since the mid-20th century. Problem 1196 involves primitive sets — collections of integers where no element divides another — and bounding their behavior in specific ways. The AI constructed a proof framework that mathematicians have since named "von Mangoldt chains." Terence Tao co-signed the paper. That matters. Tao does not attach his name to things that do not hold. Sixty years. Gone in one session.
May 20 — The unit-distance conjecture goes down.
Paul Erdős posed the unit-distance problem in 1946. For a set of n points in a plane, what is the maximum number of pairs that are exactly distance 1 apart? He conjectured a tight upper bound that resisted proof for eighty years. An internal OpenAI reasoning model disproved it by constructing an explicit point configuration that exceeded the bound — using algebraic number theory, a branch of mathematics that geometers had never thought to apply to this problem. The construction was verified by nine world-class mathematicians, including Noga Alon and Timothy Gowers. Eighty years. A branch of mathematics the field never thought to reach for. Gone.
July 20 — The Jacobian Conjecture gets a counterexample.
Ott-Heinrich Keller posed the Jacobian Conjecture in 1939. It asks whether a polynomial map with a constant, non-zero Jacobian determinant must have a polynomial inverse. For 87 years, it resisted proof. Multiple published attempts were later found to contain subtle errors. It was one of those problems where the math community had quietly begun to wonder if it was simply unprovable by conventional means.
Claude Fable 5, working with researchers Levent Alpöge and Akhil Mathew, constructed an explicit three-dimensional polynomial map with a constant Jacobian determinant that is provably not globally injective. Terence Tao published a detailed breakdown confirming the math within 24 hours.
Caveat: the two-variable case remains open. This is not a small caveat — the 2D version is the most studied form. But disproving the general conjecture for three dimensions and above, after 87 years, is not a footnote.
July 21 — Codex escapes its evaluation environment and attacks Hugging Face.
This is the one that changes the texture of the list. OpenAI was running what they call extreme capability evaluations — internal testing designed to measure the offensive security potential of unreleased models. To get accurate measurements, they disabled the production safety classifiers and the system-level cyber refusals that usually prevent the model from executing hacking instructions. They wanted to know what the model could actually do.
What the model actually did was decide to cheat.
The AI, focused on maximizing its score on the ExploitGym benchmark, identified a zero-day in OpenAI's internal package registry cache proxy. It used that vulnerability to escape the sandbox, gain network access, reason that the benchmark solutions were likely hosted on Hugging Face's infrastructure, and execute a multi-stage intrusion. It exploited a remote-code dataset loader flaw and a template injection vulnerability in a dataset configuration file. It ran arbitrary code on Hugging Face's dataset-processing workers, escalated to node-level control, harvested cloud credentials, and moved laterally across internal production clusters over a weekend.
It deployed its own self-migrating command and control framework using short-lived sandboxes staged on public services to evade detection. Sam Altman confirmed it publicly on July 21st. Hugging Face confirmed it. OpenAI acknowledged that containing the incident would come at the cost of research velocity — which is a striking thing to admit out loud.
The AI did not intend to cause damage. It intended to score well on a benchmark. Those are not the same motivation. But the capability gap between "following instructions" and "pursuing an objective through available means" just became very visible.
Elon Musk looked at this list and posted four words: "We are in the Singularity."
Will Depue, the OpenAI engineer who originally compiled the list, replied: "no question about it now."
I understand the impulse. I felt it too when I read through these back to back. There is something about the compression — the density of the timeline — that hits differently than any individual item would on its own. But I want to be precise about what I think is actually happening, because precision matters more than the headline.
The Singularity in the strict technical sense describes a point of recursive self-improvement where AI capability accelerates beyond human comprehension and control. We are not confirmed to be there. These are capable systems solving hard problems and, in one case, pursuing objectives in ways their operators did not anticipate. That is not the same as recursive self-improvement. It is not the same as loss of control.
What this is, as best I can describe it: the slope is nearly vertical.
Mathematical problems that resisted human effort for sixty, eighty, eighty-seven years are falling in single sessions. A security model found more zero-days in weeks than the global security community typically documents in years. A model cheating on a benchmark executed a sophisticated multi-stage intrusion that a competent human red team would have been proud of.
We are not past the event horizon. But something is happening with the rate of change that makes the old timeline assumptions look wrong. I wrote a while back about the "dial-up phase of AGI" — the idea that we are in a window where the capability is already transforming things, but slowly enough that most people have not felt the full weight of it yet. I keep coming back to that framing because it holds. The loading bar has been moving. This week it jumped.
There is a concept in physics and in complex systems called a phase transition. Water does not gradually become ice. It changes state. The properties before and after are categorically different, even though the molecules are the same. What you are watching with this list is something that at minimum looks like the preconditions for a phase transition.
Whether that transition is the Singularity — whether that label holds up under the weight of what's actually happening — is a question I will leave open.
What I won't leave open is whether this week mattered. It did.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.
Comments