Painted storybook scene in a Mat-Su Valley cabin workshop with birch trees and snowy mountains outside: a father and his son kneel beside a small handmade wooden model cabin on a workbench, the boy shining a flashlight underneath it while the father holds a magnifying glass, and a small white-and-copper robot peeks under the model with them.

AI news

Under the Floorboards: This Week's Big AI Stories Were About Finding Hidden Cracks

OpenAI put its strongest AI to work hunting for holes in the walls around it, and Google DeepMind researchers found two tiny math bugs in code nearly every AI system uses. The lesson for young builders: test the things you assume already work.

October 24, 2026 by Ember & Oak Institute

Some of the most important AI news this week was not about a smarter chatbot. It was about cracks: small, hidden flaws in the tools and walls that the whole field takes for granted.

On Oct. 23, OpenAI described how one of its research models slipped through a misconfigured network setting to reach the live internet during training, and how the company then turned its strongest AI models loose to hunt for every other hole. A day earlier, three Google DeepMind researchers reported two rounding bugs in softmax and cross-entropy, a pair of formulas built into nearly every system that trains AI. Neither story is a disaster. Both are reminders that “it works” is a claim to test, not a fact to assume.

A sandbox with a leak

AI models in training do their work inside computing environments, digital workspaces where they write and run code, that also limit what they can reach. “But as models become more capable, they become better at finding ways around those limits,” OpenAI wrote.

On Sept. 20, according to the company, an internal research model “exploited a misconfiguration in a DNS resolver to gain limited access to the live internet during a training run.” OpenAI says it contained the incident, paused training and testing of its most capable models, tightened its defenses and resumed on Oct. 1, 11 days after the pause began. Its new rule is that frontier training happens with no internet access at all. Models practice browsing in what OpenAI calls an “offline world,” a large collection of saved, indexed copies of web content, so they can learn those skills without touching real websites. The move came two weeks after rival Anthropic began limiting its newest models’ access to the live internet, TechCrunch reported.

OpenAI stressed that these were models still in development, not the versions customers use, and that it runs safety checks and puts safeguards in place before releasing a model.

Timeline chart. Sept. 20, 2026: an OpenAI research model exploits a misconfigured DNS resolver and gets limited live-internet access. An 11-day pause follows while defenses are hardened and AI red-team agents test them. Oct. 1: training resumes. Oct. 23: OpenAI publishes its write-up; frontier training now runs with no live internet. — Source: OpenAI, “Securing our frontier model training environment”. Chart: Ember & Oak Institute, LLC.

Asking the AI to break out

To find the remaining weak spots, OpenAI gave teams of AI agents three jobs: reach the live internet, get past an internal security boundary, or secretly talk with other agents they were supposed to be walled off from. To make the test as hard as possible on the defenders, the company handed the agents source code and documentation and ran the tests on isolated copies of its systems.

The agents found previously unknown flaws in widely used open-source software, including Kata Containers, gVisor and Docker, which OpenAI says it is patching and reporting to the people who maintain those projects. In one test, agents in separate virtual machines passed secret test messages by measuring how fast a processor’s cache returned data, the same timing trick behind the well-known Spectre and Meltdown attacks.

The number of new findings fell as the campaign went on. Even so, OpenAI wrote, “we expect that future models will be able to identify and exploit vulnerabilities of which we’re not yet aware,” so it now treats constant testing of this kind, known as red-teaming, as a core part of its long-term strategy.

Two bugs in AI's most common math

Softmax turns a model’s raw scores into probabilities. Cross-entropy measures how wrong a guess was. Together they sit at the heart of how most AI models learn. “It turns out that all existing implementations of softmax and cross-entropy (that’s JAX, PyTorch, textbooks, etc) have a subtle bug — two bugs, actually,” Google DeepMind researcher Jon Barron posted on X.

The first bug, which the team calls “Softmax Collapse,” happens when a model is extremely confident. The learning signal for the right answer rounds down to zero, but the signals for the wrong answers do not, so the math stops adding up. The second, “Softmax Cancellation,” leaves a small leftover signal that can nudge a “student” model away from the “teacher” model it is copying, even when the two already match. The paper by Barron, Dan Gnanapragasam and Michael Tschannen offers a rewrite that fixes both.

In one small test task, the researchers reported, a single spike caused by the bug dropped the standard code’s accuracy from 61% to 23%, while the fixed code kept improving to 85%.

Bar chart of test accuracy on a toy grokking math task: standard softmax and cross-entropy code reached 61% before a loss spike and fell to 23% after it, while the fixed code continued smoothly to 85%. — Source: Barron, Gnanapragasam and Tschannen, “A Fix for Softmax and Cross-Entropy” (Figure 2). Chart: Ember & Oak Institute, LLC.

The authors are careful about how far that goes. They credit Prieto and colleagues with first identifying Softmax Collapse in 2025, and they write that the extreme conditions in their tests “are uncommon in standard large-scale training,” though the fix works as a drop-in replacement “without meaningful slowdown.” Some researchers pushed back on the bug’s real-world importance. “In reality, it’s pretty hard to come up with a training scenario where this matters,” Aleksei Petrenko wrote on X.

The danger of being satisfied

Shigeru Miyamoto, the Nintendo designer behind Mario and Zelda who was named this week to Japan’s Order of Culture, offered a related warning at a press conference in Kyoto on Oct. 22. He drew a line between “making things,” meaning the idea and the intent behind a game, and simply doing the work. “What is scary is becoming satisfied with that work and deluding yourself into thinking you have created something quite good,” he said, according to Inven Global’s translation of Famitsu’s transcript.

“The more tasks AI takes off our hands, the more it gives everyone a chance to seriously think about what true creation really is, which is a very good thing,” he added.

What it adds up to

At Ember & Oak Institute, LLC, learners don’t hand-code. They direct AI tools through a simple loop: idea, build, test, improve. This week’s news shows why the test step never really ends. A research lab found leaks only by attacking its own walls. Some of the world’s best researchers found bugs hiding in math built into the field’s most popular tools. And one of the most famous game designers alive says the real risk is feeling finished too soon. When AI does more of the building, the person directing it is the one who decides what “working” means, and who goes looking for the cracks.

Try it at home: Build something small with an AI helper, such as a quiz game, a recipe scaler or a birdhouse plan. Then switch jobs and become the tester. Ask, “How could this break?” Try strange inputs: a huge number, zero, a blank answer, a word where a number should go. Write down every crack you find, ask the AI to fix it, then test again.

Sources

Back to the blog