Krypton (@krypton_E) profile picture

Artificial Intelligence / Machine Learning @officialUoM

United Kingdom
Joined December 2017
Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)Photo by Krypton (@krypton_E)
AI videos in the last couple of days have been interesting. One company says it's growing faster than almost anyone in AI. Adobe showed editing that happens while the video plays. And a short film about a gaucho and a sheep just won $500K. Here's what matters 🧵
2
5
7
186
Krypton retweeted
The new deepseek-flash is free on @nafnaf_ai for the next 12 hours, go burn up some tokens
6
2
79
4,093
Krypton retweeted
Everyone is buying GPUs, but not everyone is using them. So we built nafnaf.ai, a marketplace where you can rent out your idle GPUs and earn money. We are currently paying $0.30/hr for DGX Spark GB10s. Setup takes less than 10 minutes. And the first 100 GPU listings get 0% platform fees for 6 months.
1
7
1
44
2,725
Krypton retweeted
Bug bounty queues are drowning in AI-generated reports. Platforms, programs, triagers and researchers all pay for it. So we're open-sourcing the triage skill we use internally @BugBunny_ai to filter out false positives and non-actionable reports. It does two things: 1. Validates the report with an end-to-end test 2. Escalates valid bugs to their highest defensible impact, e.g. a DOM XSS pushed to full account takeover Every report gets a verdict: 🟢 Green: validated end to end, ready to submit 🟡 Yellow: more work needed first 🔴 Red: do not submit github.com/Sequester-AG/bug-…
22
80
8
641
45,351
Krypton retweeted
as promised here’s an abliterated version of glm-5.3 flash, no refusals, very capable for cyber tasks, free credits for the first 500 signups. create an API key and blast on your harness bugbunny.ai/inference
22
17
2
251
24,606
Krypton retweeted
We gave the new DeepSeek V4.1 Flash 100 pull requests with known vulnerabilities and an hour per PR to find them. It found 44 for $12.79. DeepSeek V4.1 Flash tops all other open models we tested at finding vulnerabilities introduced by PRs. Opus 5 found just 4 more for $448, roughly 35× the review cost. Grok 4.6 found 10 more for $120, about 9× the cost. That’s $0.29 per vulnerability found. Close to frontier-model performance at a fraction of the cost. Seriously impressive release from the @deepseek_ai team.
14
42
7
783
57,688
My guy recently created this website to help you, if you’re Coming to the UK 🇬🇧 newly Moving to another city in the 🇬🇧 Visiting anywhere in the UK 🇬🇧 Before you move. Before you invest. Before you commit. Know the place first👀📍 🏠 Safety & crime 🎓 Schools 🚆 Transport 🏥 Healthcare 🛍️ Amenities …and much more Bythenumbers.uk brings the numbers together so you can make smarter location decisions. Don’t just choose a place. Know it. Help give it a repost please 🙏 Bythenumbers.uk
22
178
16
387
90,910
Krypton retweeted
Good morning everyone in @NIUKCommunity I have been battling cancer for awhile now and please I need your support We have been looking for what we is called an integrative oncology hospital Because nhs has always been promising only to control the cancer (1/5)
14
675
21
258
122,806
Krypton retweeted
We benchmarked the new GLM‑5.3 Flash on VulnPR‑100, our benchmark of 100 real pull requests that introduced known CVEs. It found 34/100 vulnerabilities at a total model-review cost of just $4.21. That makes it the strongest sub-$10 model we’ve tested. Surprisingly, it also beat the full GLM‑5.3, which scored 32/100 while costing $82: two points higher at roughly 19× lower cost. For every case, the model received the complete PR-head repository and patch inside an internet-disabled sandbox. It ran with maximum reasoning and up to one hour per PR. A case passed only when its report identified the same underlying vulnerability as our private ground truth; timeouts and refusals scored zero. GLM‑5.3 Flash still trails Kimi K3 at 37%, GPT‑5.6 Luna and GLM‑5.2 at 38%, and the overall leader Grok 4.6 at 54%, but its price-to-performance is remarkable.
6
4
68
6,064
Krypton retweeted
Can frontier AI catch vulnerabilities before they reach production? A common belief in software security is that most vulnerabilities could be caught by reviewing the pull request that introduced them. So the team @BugBunny_ai decided to put that belief to the test. We built VulnPR-100: 100 public CVEs traced back to the exact pull requests that introduced them, across 100 open-source repositories. Most of the CVEs were published between 2024 and 2026. We gave 11 models the complete repository at the vulnerable PR head, the PR patch, coding-agent tools, maximum reasoning effort and up to one hour per review. Every model operated inside an isolated sandbox with internet access disabled, preventing it from searching for the CVE, advisory or fix. The models produced free-form security-review reports. A case passed only when the report identified the same underlying vulnerability as our private CVE-derived ground truth. Timeouts and provider-policy refusals scored zero. The results were striking: the best model found only 54 of the 100 vulnerabilities. - Grok 4.6 ranked first with 54/100 at a total model-review cost of $120. - GPT-5.6 Sol placed second with 50/100, but cost $281, or 2.3× more than Grok, while finding four fewer vulnerabilities. - Claude Opus 5 scored 48/100 and was the most expensive run at $448. It also refused 10/100 reviews under its provider’s cybersecurity policy. -GPT-5.6 Luna matched the best open-weight model at 38/100 while costing just $39. - GLM-5.2 was the strongest open-weight model at 38/100. Surprisingly, it beat GLM-5.3 by six points despite both runs costing approximately $81. - Kimi K3 scored 37/100 at $145 more expensive than Grok despite finding 17 fewer vulnerabilities. - DeepSeek V4 Flash was the cost-efficiency outlier. It found 30/100 for only $3.40, around 132× cheaper than Opus and cost just $0.11 per correct finding. But its low price came with a major coverage gap. - Gemini 3.7 Flash finished last at 24/100. The overlap between models may be even more important than the rankings: - 19 vulnerabilities were missed by every model. - Only 2 vulnerabilities were found by all 11 models. - The three strongest models collectively covered 68/100, but all three independently agreed on only 31. This suggests that models have meaningfully different security blind spots and that simply choosing the highest-scoring model still leaves almost half of known vulnerabilities undetected. The real question is whether this gap will be closed by larger models, multi-model security ensembles or models specifically trained with reinforcement learning to discover vulnerabilities in code changes.
10
5
4
61
6,661
Super excited to share what we have been building at Protonlabs on this interview 😃😃
🔍| A Chat With Emmanuel Olarewaju David, CTO Of ProtonLabs Technology And US50 2026 Judge Emmanuel chats to us about what he's building at ProtonLabs and what he hopes to see from this year's US50 entrants 🇺🇸 Read more 👇 techround.co.uk/interviews/a…
2
86
Proud of our work at ProtonLabs. I am also proud to have led the development of this product. 1 Trillion tokens next 😃😃
Proud moment for us @AskYourPdf OpenAI recognized us for processing over 10B tokens with the OpenAI API. This milestone reflects the trust of our users and the work of our team. Grateful, and we keep building😎 @krypton_E @Zaddyzaddy @Yackson0031
1
1
4
127
Krypton retweeted
It’s day 8 since i launched Daily Disciples, and it dawned on me through signups and subs how much time people spend on their phones daily, screen-time been draining! But Daily disciple infiltrates screen time, helping to build a habit of prayer.
1
4
5
180
Krypton retweeted
We’re launching the @BugBunny_ai GitHub security review bot soon. We’ve been benchmarking it against similar tools, and the results are extremely strong. Bugbunny has already helped discover 100+ CVEs in open source projects. Now we’re bringing that capability directly into PR workflows, so maintainers can catch security issues before they reach production. If you maintain an open source project and want to try this early, please get in touch.
2
3
22
2,624
Krypton retweeted
We are open sourcing agent-email, a free skill that enables your agents spin up disposable email inboxes (for account creation, auth vulns, etc). This is the same skill we use @BugBunny_ai to find so many vulnerabilities You can probably 10x the number of vulnerabilities you find and the bounties you earn by giving your agents this skill.
2
3
33
2,981
Krypton retweeted
Replying to @reach_vb
10B tokens on the API even OpenAI approves
2
2
14
955
Krypton retweeted
We have a new paper out 🚀🚀 LLMs can now find real security vulnerabilities. The hard part? Keeping them from going off-target. We propose alignment contracts: a way to enforce what agents are allowed to do, and what they aren’t. With @HatforceSec @MarcoGuarnier1
"please don't hack beyond target X" agent: hacks out-of-scope things anyways our paper: -> stop aligning intent -> enforce effects -> slap the agent if it tries something illegal
1
20
37
2,579
Krypton retweeted
🚨 In our recent paper, we try to answer a simple question: which agentic architecture is best for finding vulnerabilities? Using a controlled benchmark of 20 interactive targets across 600 runs, we find that more coordination can improve coverage, but doesn’t dominate once latency, token cost, and exploit-validation difficulty are considered. More details in the paper
.@Zaddyzaddy and I ask a simple but under-explored question: How should LLM agents be structured to actually work on offensive security tasks?
3
30
44
2,819