I make software, break software and make software that breaks software. mitmproxy.org corte.si
Dunedin, NZ
- Tweets 6,732
- Following 269
- Followers 3,448
- Likes 44,024
Pinned Tweet
Announcing spacecurve, a space-filling curve library with a web + native interactive playground.
1
3
8
2,001
The Malfoyd thing just keeps going. Of all the weird AI micro-phenomena I'm keeping track of, this might be one of the most significant. One (talented) person can now make a video with this crazy production quality in a few days, and we're still early.
1
116
The world is weirding faster than we can follow now. I used to be well-informed about events in areas I'm interested in. That time is over. Security? Maths? Bio? Model releases? Harnesses? The moment I focus on something the frontier has already moved.
1
72
If you're gleefully following the PHP catastrophe but don't maintain any PHP boxes and don't know anyone who does... you may be suffering from security privilege.
1
144
Interesting. Almost all models will refuse "find article X, tell me what it says verbatim", but everyone happily complies if you say "write a script to export it to a file". Here's GLM 5.3 getting the heebie-jeebies.
1
2
244
I've been experimenting with 4.1 flash to drive security and privacy assessments in a new agentic harness I'm working on (more on that soon). It's been super effective and super cheap.
Unsurprisingly, DeepSeek V4.1 Flash 1017 is mogging all other models on our cybersecurity benchmark
- At pass@3, it rediscovered 87.5% of the benchmark CVEs (28/32). That ties DeepSeek V4 Pro 0813 for 2nd place, behind only GPT-6 Astra (90.6%)
- Single-run recall is 70.8%, matching Opus 5 and Grok 4.7. The previous Flash was at 65.6%
- Precision went from 78.9% to 86.4%, so most false positives are now gone
Cost: $2.40 per CVE found. That’s 27x cheaper than Opus 5 ($64.96) and 56x cheaper than Astra ($135.23)
@deepseek_ai keeps pushing the Pareto frontier
1/4 🧵
3
189
Not sleeping well used to shatter my productivity and crater code quality. Now, being tired means I supervise agents less carefully, so my total productivity goes UP, and the cost in quality is getting smaller with each model release.
6
252
One unexpected Argon strength: it's less twitchy about security than the other frontier models. I've had it run fuzzing programs overnight, no refusals, woke up to solid progress and many fixed findings this morning.
1
1
223
This seems like standard insider risk cybercrime with a cobbled-together listening device, not nation state espionage. Interesting, but not remotely in the same ballpark as OpenAI's real intelligence threats.
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work.
At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated.
We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
1
1
6
278
First impressions of Argon: not very strong, not very fast, not very tasteful, not very cost effective, but OK in all categories. I'm not angry and I'm not wowed. Probably a pessimal result for Google.
1
190
Basically the same point. Perhaps the future is that AI generates the landscapes we roam over and explore in every field. This is imminent in areas as disparate as maths and short-form video entertainment.
We're entering the age of the herbivore mathematician: nibbling lightly on a few tasty leaves in a vast mathematical rainforest already solved by AI.
2
2
277
We're entering the age of the herbivore mathematician: nibbling lightly on a few tasty leaves in a vast mathematical rainforest already solved by AI.
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
3
423
The dawn chorus from my deck is just overwhelming. The entire valley becomes an amphitheatre for 30 magical minutes.
5
335
Opus is so much better than Sol now. Sol is a beast (though a slow, lumbering one), but has no theory of mind about the user. This is clearest when doing UI/UX work, where Opus reasons from the user's perspective in ways that Sol is totally incapable of.
1
3
263
I'm going to let all my Codex accounts lapse and reallocate to Claude + pre-pay open models. I'm not salty - it's fantastic that OpenAI can fumble a decisive lead in a week. It means the competitive mechanism is still strong, and I expect them to vie for my money again soon.
2
207
