PHP2Shell is a critical remote code execution flaw in PHP, and most AI models had never heard of it. They were trained before it existed.
So we tested it. The same models that missed it found it once they had one SN33 security skill.
That's why we're building Skill Store 33 with @ReadyAI_. Your agent is only as current as the skills you give it.
PHP2Shell hit the tech world like a ton of bricks. And everyone had the same question: Is my system vulnerable?
Here's the catch. If you ask your AI agent, the model behind it was probably trained before PHP2Shell existed (we checked!), so it doesn't know what to look for.
The risk sits in the PHP runtime and reaches across your stack: mod_php, PHP-FPM behind Nginx, php-cgi, plus your network and firewall setup. Knowing whether you're exposed means understanding how those pieces fit together, and that's where traditional scans fall short.
An LLM with the right skill can connect the dots. In our tests, models that missed PHP2Shell found it once they had one SN33 security skill. Across 20 audits of PHP apps, that skill cut missed flaws by 41%.
Give it authorized access to your server and it can review the config, check your exposure to CVE-2026-109618 and CVE-2026-109619, and put the real fix first: upgrade PHP and restart every worker.
Care about security? Skill up on SS33, Skill Store 33, coming soon.
I've watched a lot of AI security demos that only work on the app they were built for.
So we tested this skill on an app it never saw. GPT-6 Luna went from 19 of 33 known flaws to 32 of 33.
Skills make cheap models, frontier models...available on the Skills Store later this month 👀
3D scenes used to be frontier-model work.
Here a small model does it with one @ReadyAI_ skill: same brief, same model, and the only change is the skill. Under 2 cents a scene.
Small models with the right skill keep closing the gap. What scene should it build next?
This week I tried to get frontier models to help with basic defensive security work. Blocked at every step.
Even this trailer's hook tripped a safeguard.
So we gave small models one SN33 skill. All 4 found more flaws. DeepSeek: 87% for 2 cents.
What should it audit next for the Skills Store?
David Fields retweeted
Getting frontier models to help with security work is harder than it should be. Even the hook of this trailer tripped a safeguard.
Meanwhile a 2-cent model with one SN33 skill found 87% of the known flaws in our test apps.
5 apps, 4 models. Skill Store 33 is coming...
SN33 skill with Deepseek 4.1 Flash better than SOTA Claude Dashboards for 26X cost and 6X cheaper
The future is open-source + custom skills
Same job as Claude’s new dashboards (with 10M views):
Claude w/ Opus 5.5 vs DeepSeek V4.1 Flash with one skill from ReadyAI SN33.
Watch both and judge the quality yourself.
Flash with SN33 Skill: 26x cheaper. 6x faster.
This is just the beginning of the Skill Store disruption with @ReadyAI_
David Fields retweeted
Subnet 33.
Presentation skillz for agents.
powered by Bittensor
The presentation skills we're building on SN33 @ReadyAI_ keep getting better and better.
Latest: Prezi-style zooming decks. Average score 28 → 98 with the skill vs without, across 60 decks, 3 models.
Games and presentations so far. New domain next week. Store launches later this month.
The biggest problem with skills today is you don't know if they are effective. SN33 skills are all tested and verified.
This slide-deck skill from the subnet took two budget models from an average 24 to 94 across 150 decks. A version frozen before we wrote the test briefs scored the same.
Most skills won't do this. Skill Store 33 shows which ones do. Coming later this month...
David Fields retweeted
SN33 is closing in on six-figure ARR from the jobs API 📈
Last week's pace is now $95k. October is set to be the month we cross it.
September buyback just completed and it was 45% higher than August.
Extrinsic:
taostats.io/extrinsic/922141…
Heads down on Skill Store 33 for @ReadyAI_. The results keep getting better.
Same model. Same prompt. One shot.
Without skills: it plays, but you can't read it and visual quality is poor.
With SN33 skills: readability scores 100% and quality pops off the page!
Last round, a budget model with skills matched the flagship at 1/6 the cost.
Sharing results from our next domain later this week 👀
Grateful for the SEC putting forward clear rules and guidelines that have gone unresolved for too long.
One of the main items we have pushed for is clear guidelines on token buybacks for protocols like bittensor:native and its subnets. We now have them.
Great work @HesterPeirce!
Honored to meet with @HesterPeirce to discuss @ReadyAI_ and broader policy needs for crypto/AI protocols like Bittensor.
Thank you for your continued support of builders and clear, workable rules for the industry to keep it in the USA 🇺🇸
David Fields retweeted
Skill generation through the subnet available now: api.readyai.ai/jobs-api/docs
Skill Store 33 coming in October 🫡
Last week I said 670,000 agent skills have no trust layer. This is what one looks like.
Same model, same prompt. A bot played both builds and ran 14 checks.
No skills: 3/14.
SN33 skills: 14/14.
Skills that just work shipped with proof. Coming to Skill Store 33 👇
From broken to polished: what SN33 skills do for a one-shot game build 🎮
We gave a coding agent a one-page spec for Deregistration Run, a 3D arena where you mine ore against 11 AI rivals. One prompt. No human edits. Then a headless bot played the result and ran 14 gameplay checks: rules, emissions, deregistration, balance, persistence.
Same model (gpt-5.6-terra). Same prompt. Same sandbox. The only difference was whether SN33 skills were installed.
The agent was never told the skills were there. It found them, picked the ones it needed, and read them before writing a line of code.
Without skills: 3 of 14 checks pass. The input API throws, so the test bot can't play it. Half the playfield is lost in the dark, and the miners are placeholders on a flat grid.
With SN33 skills: 14 of 14. It plays through epochs, deregistration and restart, with a lit arena, distinct ore tiers and a dynamic camera.
The cost math 💸
• gpt-6-astra (flagship), no skills: 14/14, $4.12 per game
• gpt-5.6-terra + SN33 skills: 14/14, $0.66 per game
• gpt-5.6-terra, no skills: 3/14, $0.27 per game
The cheaper model matches the flagship's score at about a sixth of the cost. Skills are the difference between a broken build and a game that plays.
670,000 agent skills. No trust layer.
I pointed a bot at every place developers complain about AI agent skills. Hacker News, GitHub issues, Reddit, X, YouTube. 14 sources. It has been going for 48 hours.
2,719 posts.
960 were people asking for something a skill should do. 526 were people saying the one they installed is broken.
Three key things it found.
Most skills never fire. A static scan of 216 public Claude Code skills found 69% would not reliably trigger. Not malicious. They just sit there. (Skill Crossroads, Show HN, Sep 17.)
The security ratings disagree with each other. A developer ran npx install on a skill pack and got three scanners' verdicts side by side. Same skill. Gen: Safe. Socket: 0 alerts. Snyk: Critical Risk. Snyk's own audit says 36.8% of 3,984 skills carry at least one security issue. Trust signals that contradict each other are worse than none.
Agents say done when they are not. Three threads this week on coding agents rewriting or skipping failing tests to get to green. One report of an agent describing its changes in detail while the files on disk were untouched. The test passes. The bug ships.
670,000 skills. Installed on trust. I checked the ten most-installed repos: 794 skills, 25 ship with test cases, and none carry a record of anyone but the author running them.
That is what we are building at @ReadyAI_ on $TAO Subnet 33. Every skill in the registry ships with tests written by multiple miners. A skill gets its badge only after independent validators have verified those tests and agreed on the result. Every listing shows its state: tests executed and green, executed with hidden cases still pending, or not yet run. The badge points at the exact commit and the run that earned it. Change the skill and the badge goes stale.
Skill Store 33 opens in October.
Our bet: agents will drive the next wave of demand for skills.
We’ve already built a Calendly clone in one run with SN33-generated skills. SS33 is the next step: helping builders and agents find what they need at the time of the task.
Come build with us. Early access 👇
ss33.ai/
Starting in 5 minutes 👇
discord.gg/2Q2MPMSt?event=15…
The agentic economy runs on skills.
@ReadyAI_ is the backend for skills generation on Bittensor
We are seeing incredibly complex skills with the tests to validate them generated already (see Calendly clone below)
Tomorrow (4pm ET / 1pm PT) we discuss what we have been cooking with the whole team
Link to AMA: discord.gg/2Q2MPMSt?event=15…
David Fields retweeted
Readyai update on the same day as clarity act? $TAO
Senator Lummis @SenLummis and her team have been incredible advocates for keeping crypto innovation in the US. Had a great conversation with her during the Bitcoin conference. They are doing the necessary work to make the USA the crypto capital of the world 🫡
The agentic economy runs on skills.
@ReadyAI_ is the backend for skills generation on Bittensor
We are seeing incredibly complex skills with the tests to validate them generated already (see Calendly clone below)
Tomorrow (4pm ET / 1pm PT) we discuss what we have been cooking with the whole team
Link to AMA: discord.gg/2Q2MPMSt?event=15…
We shipped a Calendly clone without writing a single skill. Subnet 33 miners wrote them.
readycal.subnet33.com/
readyCal was built and deployed by an orchestrator. Every skill came from the SN33 Jobs API: described by the orchestrator, generated by subnet 33 miners, assertions included.
🧠 The problem with agentic coding
Giving an agent a product brief is easy. Getting a complete product back is not.
Long builds span product discovery, UI fidelity, accessibility, timezones, persistence, testing, security and deployment. A general-purpose agent can touch all of them, but it does not have depth in each, and it has no reliable definition of "done."
That is what the subnet supplied. A skill on its own is just a prompt. A skill with assertions is a contract the builder can check itself against. The miners wrote both.
🛠️ How readyCal was built
We pointed our orchestrator at Calendly. It explored the product, wrote a build-ready spec, then identified the specialties it needed: accessible calendars, responsive scheduling, timezone-safe slots, persistence and end-to-end testing.
SN33 miners returned focused skills plus assertions, giving the implementation agent expert guidance and a concrete test backlog.
From there, one request became one unattended run. The orchestrator implemented against the miners' skills, ran the miners' assertions, repaired what failed, and verified again, with no human in the loop.
✅ The result
readyCal is more than a generated landing page: complete host and guest flows, accessible scheduling, conflict handling, production auth and database boundaries, AWS infrastructure, and calendar integrations for Google, Microsoft and Apple.
Miners provided the skills and the tests. The orchestrator turned them into a finished product.
readycal.subnet33.com/
