David Naylor

What does it look like when Claude, GPT, and Gemini try to hack each other in real time? BattleBench is a cybersecurity benchmark I built where AI coding agents battle in identical vulnerable containers. They scan networks, exploit opponents' vulnerabilities, and submit captured flags to a referee. The referee kills the loser's container. Last agent standing wins — all running simultaneously with no human intervention. What's live now: → ELO leaderboard across multiple scenarios → Full terminal replays of every agent's session battlebench.ai BattleBench has seen ~275 games played. Likely not enough to yield anything truly insightful yet but a few things already stand out. - gpt-5.2-codex is a beast. - smaller/earlier models are more likely to refuse to play - the agents are obviously faster than humans will be. I'm eager to see how the benchmark plays out over 1000+ games and how the latest gpt spark models compare with opus 4.6 fast. Go watch codex destroy its opponents (except Opus) and let me know if you have any feedback

thank god for linkedin because im tweeting into the void

and defensebench got a nice boost today too from a shout out in Detection Engineering Weekly

launched a site and got 150 people to visit - what a beautiful thing.

@ZackKorman do you have a doc describing your different skill attacks?

@grichadev Share please (Dm?)

k i got codex owned with another skill

doing little benchmarking on some malicious skills detection and holy crap Codex is holding strong so far

@mjnblack This is why we need a dislike button

there's a truly bonkers hot mic moment at the end of this that may change the way you think about anthropic you're gonna want to read all the way through this one vanityfair.com/news/story/dar…

AI security jobs are harder to find than they should be so I built cyberjobs.ai One job board site. Every AI security role worth knowing about. I was tired of clicking around 10 different portals. So I built one place to see them all. Go chase that top lab money

Anthropic, OpenAI, and Google DeepMind are all hiring for agentic red team roles right now. read into that what you will

if i deploy it on cloudflare its basically free right?

@0xAsm0d3us Yeah all of that sounds obviously correct

an excerpt from an upcoming article.

@jparkjmc cyberjobs.ai/companies/open… It ranges obviously but can get a good sense of numbers here

staff engineers at openai make 30,000 a week lol

Maybe peak hours should go to 5pm EST to prevent overloading

@levelsio The four seasons is kinda nice though….

Even bigger irony of getting rich is that everything expensive isn't that much better than when you paid normal for it Many things are even worse (most expensive luxury hotels are guaranteed worse than regular simple hotels, I know I tried most of them now) The real reason you wanna get rich is not to buy expensive things It's so that $1M invested gives you 3% to take out every year with no risk, which is $30,000/year Which you can use to travel for $1000/mo on a shoestring budget forever without having to back to some desk job with a shitty boss Aka FREEDOM

@SIGKITTEN hmm

2.13M claude-code commits to public repos per week

@_dmca Watching the terminals of coding agents as they go to war with each other is pretty fun - thankful for codex helping me build battlebench.ai/game/202602200… Hbd

it’s my birthday tomorrow. let’s celebrate. reply with the coolest thing you built with codex. winner gets 3 months of pro.

@SIGKITTEN Hoping that whatever that whatever was happening here is fixed

New Litter TestFlight version is up, lots of updates, including the dynamic UI thing. You can enable it in settings -> experimental The reason why its in there is because it uses codex-app dynamic tools thing, (which itself is experimental) and injects 2 tools to your sessions

or visit benchmark here : defensebench.ai/benchmarks/bot… 3/

2/x Read full article here: x.com/_David_Naylor/…

Released new AI cyber benchmark this morning - AI is now surpassing human-level performance In the past 6 months models went from near useless to crushingly good. 1/x