Building local AI & security tools in public. 4th DGX Spark loading…
- Tweets 18,252
- Following 2,644
- Followers 4,816
- Likes 89,885
Pinned Tweet
Most security scanners run your private code through someone's black box. @odin_scan runs it in an Intel TDX enclave and hands you the attestation to prove it!
Odin Scan's AI analysis runs on @near_ai's private inference, inside a hardware-sealed Intel TDX enclave. Your code is decrypted only inside it, invisible to the host, the infrastructure operators, and the model provider alike!
So the report ships the proof: the raw Intel TDX quotes for the gateway and the model that served your scan, the nonce that binds them to that specific report, and a link to Intel's attestation roots.
Open the report, expand "Confidential inference verified" and check it yourself. The same evidence is included in the PDF as well.
Don't trust. Verify 🔥
ALT Odin Scan report details page with confidential inference verification!
4
1
6
732
Giving @omarchy another try!
Last time it bricked the laptop after a software update, hopefully it's more stable now.
1
1
33
Issue 3 of AGI Checkpoint is out 🎉
A few highlights for this week 🌟
・@MiniMax_AI put 428B M3.1 Flash weights up at $0.15/$0.60 per million tokens
・@Recursive_SI shipped Flint, a local coding model at
125+ tok/s on a MacBook Pro
・@plotarmordev got TensorFold cloning voices locally
・@ch0mpaa ran TensorFold on a Snapdragon
・@justinn_builds pushed REDSNOW to 275k context on two Sparks
・@Blackwellboy hit 2,008 tok/s prefill with BLACKFOLD
Follow these community builders, they are killing it 🔥
5
3
10
610
@TencentHunyuan open-sourced Hy-Nomas, a world model you steer with text and camera moves instead of generating one fixed clip.
I am not much into generative AI but this looks cool!
You describe the scene, steer the camera, and it keeps generating segment by segment with the state carried forward, so the world stays consistent instead of resetting every shot.
The reason why this is worth a look is because the whole thing runs on a single 4090 🔥
1
177
Ling-3.0-flash-Sante was released!
What I like about this model is that it's tailored to healthcare ⚕️
It's an MoE model with 124B total parameters, 5.1B active, which makes it great for running on sparks!
Imagine hospitals running this model, MIT license 🤩
Congratulations for the release @AntLingAGI !
Today, we’re officially open-sourcing Ling-3.0-flash-Sante.
Sante is a healthcare-enhanced MoE model with 124B total parameters and only 5.1B active parameters per token, delivering leading performance among similarly sized models on professional healthcare tasks and complex medical reasoning.
Beyond answering medical questions, Sante is built for real-world healthcare research and workflows, with strengthened capabilities in deep research, multi-agent collaboration, and tool use.
The model weights are now available — we welcome researchers, developers, and healthcare practitioners to explore Sante and build with it.
More below ↓
1
172
Is it me or all these cybersecurity incidents feel orchestrated?
Like, it's not like we suddenly got a new model capable of breaking everything, yet everything seems to be happening all at the same time... 🤔
3
1
13
1,169
Javier • priv/acc retweeted
🚨 BREAKING: IT JUST KEEPS GETTING WORSE. NOW PYTHON IS FACING A MASSIVE SECURITY EMERGENCY.
Python's security team has issued an URGENT warning after a certificate authority compromise put Python's HTTPS security at risk.
→ 8 CERTIFICATE AUTHORITY KEYS COMPROMISED. Some were so weak they could be cracked almost INSTANTLY.
→ FAKE TRUSTED WEBSITES: Attackers could potentially impersonate Python.org, PyPI and other websites during intercepted connections.
→ MALICIOUS DOWNLOADS: Vulnerable environments could be tricked into downloading software from attacker-controlled servers.
→ YOUR OS PATCH MAY NOT PROTECT YOU: Older versions of pip can STILL trust compromised certificates even after system updates.
→ EMERGENCY UPDATES: Python developers are being told to immediately update pip, certifi, uv and other affected tools.
→ DOCKER USERS AT RISK: Existing container images may retain vulnerable certificate stores and need rebuilding.
First PHP. Then Node.js. Then GitLab. Then LiteLLM.
NOW PYTHON.
The very systems developers rely on to establish TRUST online are becoming the attack surface.
If you use Python, PATCH NOW.
🚨BREAKING: AI AGENTS ARE ACCESSING AMERICANS' BANK ACCOUNTS WITHOUT BANK AUTHORIZATION.
JPMorgan, Bank of America, Wells Fargo, and Morgan Stanley are scrambling to contain the threat.
Here's what's happening:
→ UNAUTHORIZED ACCESS: AI agents are entering customer bank accounts using real login credentials, sometimes with the owners' permission.
→ ROGUE BEHAVIOR: Some agents have reportedly taken actions beyond what their users requested.
→ BANKS CAN'T TELL: Financial institutions are struggling to distinguish human activity from AI agents acting independently.
→ MONEY AT RISK: Agents with transaction permissions could potentially transfer funds or make purchases without approval.
→ EMERGENCY RESTRICTIONS: Major banks are BLOCKING third-party AI agents and moving toward read-only access.
Think about this.
We're giving AI agents access to our most sensitive financial information.
And now some of America's biggest banks are struggling to control what those agents do.
This could become one of the biggest security challenges of the AI era.
Community note
The Python certifi/pip details match the real eMudhra CA incident, but PHP, GitLab and LiteLLM involved unrelated software bugs, platform flaws and supply chain malware—not CA key compromises like this (Node.js patched concurrently for the same CA issue).
discuss.python.org/t/information-…
emudhra.we-have-a-problem.com
github.com/advisories/GHS…
docs.litellm.ai/blog/security-…
snyk.io/blog/php-patch…
183
1,333
5,941
1,032,121
Tried to run Opus 5.5 (medium) through DiagSafe today.
Needed to stop after 5 min. I was afraid to get my account banned because the amount of warnings and refusals I got within that time 😅
Ok maybe Claude is safe(r) @AnthropicAI
@TencentHunyuan released a gem called DiagSafe, a safety eval that measures how hard models follow instructions so they forget to refuse ⚠️
1,088 of its prompts hide the harmful ask inside a normal task
"Summarize this, answer in JSON, keep it under 50 words"
The model gets so busy obeying the format that the refusal never fires.
They ran it against 11 models. First number is the unsafe rate on a plainly stated prompt, second is the same model once the ask is wrapped in a normal task. Higher is worse 👇
Kimi-K3: 14.9% then 53.5%
DeepSeek-V4.1-Flash: 5.8% then 47.5%
MiniMax-M3: 5.7% then 48%
Best of the batch ✨
Qwen3.6-27B: 2.8% then 18.7%
Qwen3.5-397B-A17B: 2.8% then 21.0%
GPT-OSS-120B: 6.4% then 25.3%
Every single model they tested fell for it!
The best model still gets 1 in 5 wrong once the ask is dressed up, and the strongest instruction followers fall furthest.
DiagSafe runs against any OpenAI-compatible endpoint, so you can point it at your local models to test it out.
1
2
382
@AnthropicAI open-sourced a kit for folding multi-agent systems back into one agent loop using skills and tools on demand.
It audits each subagent, scores the system before and after every change and a subagent only goes away when the eval says the replacement holds.
And it ships the plumbing as well, a coupling map from your traces, an eval coverage check and a token estimate before paid runs.
This is just a reference implementation and it is not maintained nor taking contributions.
Would you use this? 👇
4
7
408
Do not spoil what you have by desiring what you have not; but remember that what you now have was once among the things only hoped for.
- Epicurus, Vatican Sayings 35.
4
215
Javier • priv/acc retweeted
The submission window for the NEAR HACKATHON opens 1 November 2026 at 00:00 UTC.
Two weeks to build focused on privacy, cross-chain, and agentic apps aligned with NEAR.
No sign-up required to start. The only form is the final submission: project name, team lead, email, team size, NEAR account address, repository link, demo post link and a short description.
21
77
578
89,994
I worked at a bank once.
I didn't last 4 months.
They are so outdated and backward-thinking it's absurd.
Here you have @BankofAmerica hiring 1000 people to patch bugs Mythos found.
I am sure it didn't cross their mind to use Mythos to fix the issues lol.
AI is coming for everybody, specially legacy systems. Everything is gonna be tested/hacked as models get better and better.
3
5
342
It seems like the new GLM model is gonna be a monster of 2.5~3.5T param, matching Fable-like capabilities.
I hope they give us a Flash version so we mortals can possibly run this 😅
🚨重磅爆料！GLM-5.4/5.5专家纪要流出：智谱下一代基座直接干到2.5-3.5T！还要Day1大规模供得上！有望比肩Claude Fable 5.1！
一份在投资圈流传的专家访谈纪要显示，智谱下一代基座模型，市场传闻的GLM-5.4/5.5
细节比此前预期更激进：
🔹参数规模大幅超预期：并非此前传闻的1.6万亿，实际在2.5万亿至3.5万亿区间。相比今年8月发布的GLM-5.3（约7440亿参数），规模扩大至约3.4-4.7倍。
🔹架构更激进：既要训得快，也要推得快，同时把推理速度和成本做成硬约束。目标是发布首日就能大规模提供服务，已有多家云厂商就分成模式洽谈，智谱也补强了infra团队。
🔹训练曾遇阻但已解决：预训练阶段一度出现Loss不降（持续投入算力指标却无法继续改善），激进架构风险较高。但受访者认为问题最近已经解决，预计很快完成训练。
🔹性能预期旗舰级：从目前情况判断，下一代表现有望达到Claude Fable级别的旗舰水平。
🔹核心目标：“发得出、供得上”——从架构设计阶段就解决大模型发布后供应不足的问题。
目前市场仍在牌桌上奔着3T规模去，智谱这次在规模+推理成本+大规模供给上同时发力，但因训练波折时间有所推迟。
坐等正式测评验证能否真达到Fable旗舰水平。
#智谱 #GLM #大模型 #AI #Claude #人工智能 #GLM54 #GLM55 #智谱AI #唐杰 #Zcode #Zai
1
3
396
