Pinned Tweet
Oracle patched a Java flaw Tuesday and CISA said patch now. The finder scores it 9.2 and says AI helped find it in an hour and exploit it in two more, for under $200. Oracle says 8.1, high complexity. Which number does a defender act on?
Reg Saddler retweeted
🚨 Today, Nov 18, at 06:59:34 we saw the first instance of ARTEX and Java4Shell being used from the same IP address. This indicates threat actors are combining new vulnerabilities to carry out attacks.
Reg Saddler retweeted
Genie 4 is now available in Project Genie.
It brings native audio, real-time editing, and a new agent to help you create interactive, virtual worlds.
Today’s coding agents stop when code ships. Stitch Loop, a new experiment from @GoogleLabs, keeps working after launch.
Using Gemini 4 Carbon and CodeMender, it watches live metrics, tests fixes in a sandbox, and opens PRs for you to approve — all tied to your objectives. Join the waitlist at stitch.google.com/loop.
Introducing Stitch Loop, a new Stitch experiment.
Stitch helps you design your product. Stitch Loop helps you keep improving it once it's live: a product agent built for outcomes, not just PRs.
Stitch Loop monitors your production data and lets you build loops that work continuously until your goals are met.
You can define your own custom loops, or start improving your product on day one with built in Loops for:
🛡️ Security: built with Gemini 4 Carbon and @GoogleDeepMind CodeMender
⌚️ Latency: to keep your product fast
👋Accessibility: so it works for everyone
Join the waitlist: stitch.google.com/loop/waitl…
Learn more 🧵👇
Reg Saddler retweeted
🚨 Gemini 4 Neon just hit Google Cloud's public billing catalog across 135 SKUs — 90 on Vertex AI and 45 on Gemini API:
→ Online text: $3.00 input / $15.00 output per 1M
→ Cached input: $0.15 per 1M
→ Batch & Flex: $1.50 / $7.50 per 1M
→ Priority: $5.40 / $27.00 per 1M
→ Modalities: text, image, video, and audio
Seems like a new tier 👀👀
Reg Saddler retweeted
Google's Aura: they were literally handing out Gemini 3.1 Pro for distillation
Anthropic and OpenAI cried about it, so Google had to pull it
Gemini 4 Carbon is the best model in the world for me. It sounds original and thinks in a way no other model does
You will notice that Gemini Carbon thinks differently than Claude and GPT. It often comes up with entirely different but more clever solutions than competitor models.
This is precisely because we never distilled other models. Distilling other models creates a poor imitation rather than unique intelligence.
Reg Saddler retweeted
New Harvard paper finds that 94.8% of a coding agent's tokens were re-reads of cached context, i.e. 94.8% of the text the AI processed was earlier conversation it was re-reading, not new work.
And that re-reading raises its estimated energy use about 25 times, from roughly 10 to 250 kilowatt-hours over 10 weeks.
Energy estimates for LLMs usually price single request-and-response calls. Agents re-read a growing context on every call and hand work to subagents, which those estimates can miss.
The authors logged 10 weeks of real Claude Code use by 1 engineer, about 1 billion tokens across 42 sessions. They built an energy estimator that charges cached tokens for attention reads but not full reprocessing.
About 260 cached tokens were read per generated token, and context grew 5.3× over a session. The cache-aware estimate came to about 250 kWh, against about 10 kWh from the output-based EcoLogits estimator.
If you track agent cost or emissions, count cached reads and subagent calls, and start new tasks in fresh sessions when possible.
OpenAI just published its 𝗳𝘂𝗹𝗹 𝗴𝘂𝗶𝗱𝗲 to using GPT-6.2 in Codex. You can ask GPT-6.2 Astra to use it to review your Codex setup.
Half of X is trashing 6.2, the other half says OpenAI is back. Let them fight. Tuned right, it 𝗰𝗼𝘀𝘁𝘀 𝘄𝗮𝘆 𝗹𝗲𝘀𝘀 𝘁𝗵𝗮𝗻 𝘁𝗵𝗲 𝗹𝗮𝘀𝘁 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻. 6.2 Astra is 20% cheaper than GPT-6 Astra, and 6.2 Sol costs a quarter of what Astra does. Plus also gets more messages every 5 hours than the last generation: 17–180 for Sol, 6–50 for Astra.
Don't start by maxing out Astra. Sol is enough for everyday coding. Only switch to Astra when you're tracing one bug across several services.
Keep Opus 5.5 as your main builder in Claude Code and hand the reviews to GPT-6.2 Sol. Install Herdr (brew install herdr) and its Claude Code skill (npx skills add herdrdev/herdr --skill herdr -g), run herdr in your project, open claude --model claude-opus-5-5 in one pane and codex -m gpt-6.2-sol in the pane next to it, and tell Opus the Codex pane is its reviewer: after every step, it sends Sol the diff through Herdr. 𝗔𝗻 𝗲𝘅𝘁𝗿𝗮 𝗿𝗲𝘃𝗶𝗲𝘄𝗲𝗿 𝗳𝗼𝗿 𝗮 𝗾𝘂𝗮𝗿𝘁𝗲𝗿 𝗼𝗳 𝗔𝘀𝘁𝗿𝗮'𝘀 𝗽𝗿𝗶𝗰𝗲. Why not?
A one-time config check is the right place to spend one Astra message. Start Codex with codex --search (by default it only searches a cache, so it may not see yesterday's new pages), then send GPT-6.2 Astra this prompt 👇
"Read these two pages. If you can't get the actual text, stop and tell me:
developers.openai.com/codex/…
learn.chatgpt.com/guides/pro…
Then review my Codex setup against them: ~/.codex/config.toml, ~/.codex/AGENTS.md, and the AGENTS.md files and skills in my project. Find three kinds of problems:
1. A model or effort level that doesn't fit the work. Flag Astra as my default model, or a default effort above Extra High. Use the guides: Sol for everyday coding, Medium for a quick survey, High for tracing one cause across several systems, Extra High only when several causes interact.
2. Parallel work with no instructions. For work that can run in parallel, check whether the matching skill or AGENTS.md says to use subagents, and when.
3. Too much in AGENTS.md. Flag anything Codex can work out from the code, tests and docs, and routine decisions made on its behalf.
For each issue, quote my setting and the matching line from the guide, explain what it costs and where it will go wrong, and suggest the smallest change, written as config I can paste.
Then look at this project's code and recent commits and give me a table: which kinds of work here fit Sol or Astra, and at which effort level.
Last, tell me where to check my Codex usage and how to run one task from my backlog on both Sol and Astra to compare what each one uses.
Show me the recommendations first. Don't change anything yet."
Reg Saddler retweeted
Logan lit the Gemini signal last night. And it looks like we are getting Gemini 4 Neon this morning, probably about four hours from now. Google are shipping at an incredible rate.
Reg Saddler retweeted
Media outlets need to interview technical experts instead of publishing explanations from Wall Street and the NSA that don’t make any sense. Neither is known for telling the truth.
My recent post about the outages prompted a few investigative journalists and tech vets to request my technical analysis. I’ll work on that today.
The available technical evidence suggests these weren’t cyberattacks or software updates that went wrong. It looks like a psyop, and I’ll look at the evidence to explain why.
I’ll investigate which organisations were affected, what their failures have in common, who had access to cause them, which explanations are least credible, who could have coordinated them and who benefits.
The US aviation outage seems easier to explain, and I’m not convinced it’s connected. But I’ll take a closer look.
AI giants and the US and Israeli governments, along with their intelligence agencies like the NSA, should be seen as a single connected network. They share financial and surveillance interests. They’re one big team. When you look at OpenAI and Anthropic through that lens and assume the worst of their intentions, anything, and I mean anything, is possible.
MI5 is promoting the adoption of OpenAI and Anthropic through Instagram in the name of national security, so it clearly has an interest in their success. But they’re just kids in the intelligence playground.
Jensen’s removal from Nvidia could also be connected because it serves the interests of the team I just mentioned. The last thing they want is open source models to get in their way.
Here’s another thought experiment…
Imagine a government wanted to conceal activities after an election but before a change in political control. Infrastructure outages would provide an explanation for missing communications, deleted records, financial transfers or disruptions to the movement of people and sensitive materials.
The NSA’s PR stunt around the classified communications outage is the most obvious outlier. The world’s most secretive intelligence agency, which withholds sensitive information even from the CIA and FBI, suddenly has 6 sources disclosing an internal comms failure to CNN. Another source tells the Wall Street Journal that maintenance caused it. And the public is supposed to believe they risked life in prison for that.
We’re supposed to believe that multiple tech giants independently discovered vulnerabilities, updated critical infrastructure and experienced major failures because they didn’t test and release updates properly, all at the same time. I don’t believe it.
It would be easier to believe that a hundred banks went offline because they’re more likely to share some of the same systems and services.
My original post:
teapawt.net/paul__walsh/status/212…
More to come.
The NSA is one of the most secretive intelligence agencies on the planet. It doesn't even share some of its most sensitive information with the CIA or FBI. Yet suddenly, 6 sources are telling CNN that its highly classified internal communications network is down, while another source tells the Wall Street Journal that regular software maintenance caused the failure.
I don't believe such sensitive information becomes public by accident without serious consequences for whoever disclosed it.
Meanwhile, Microsoft, Apple, US aviation communications and other critical services have experienced major outages within days of each other. I've worked in tech for 30 years and have never witnessed this pattern across so many major providers.
I'm starting to suspect something much more sinister than unrelated software failures and rouge AI.
One possibility is that major changes are taking place across critical internet and communications infrastructure. Another is that governments are preparing to use these failures to justify compulsory surveillance, tighter control over digital services and new cybersecurity or national security legislation.
Both scenarios deserve serious attention because governments and US tech companies have massive commercial and intelligence interests in controlling the infrastructure that billions of people use. And they only have to work with a handful of tech companies to achieve it on a global scale.
The NSA's decision to disclose such a sensitive failure is what I find most suspicious. If the world's most secretive intelligence agency allows the public to learn that its classified communications network has failed, I want to know exactly who authorised that disclosure and what they expect to gain.
My suspicion is that we're being prepared to accept something that governments would otherwise struggle to justify. I don't believe the public should accept another expansion of surveillance powers simply because officials claim our digital infrastructure is unsafe.
Reg Saddler retweeted
Introducing Poseidon: a state of the art cybersecurity specific model, designed for vulnerability research, finding, prove, and fix vulnerabilities in real software.
Poseidon preforms:
- 17.6% higher than GPT-6 Astra
- 2.9% higher than Claude Fable 5.2
(CWE-Bench)
Reg Saddler retweeted
AI cybersecurity startups are about to see a flood of VC money
Hackers are now using LLMs to continuously scan software updates, spot newly fixed bugs, and build attacks before teams can apply the fix — all without human involvement. That’s according to the latest report from the Google Threat Intelligence Group.
We're helping defenders counter automated threats with Google AI Threat Defense, which combines Gemini and other frontier models to find and fix flaws before attackers can exploit them.
Reg Saddler retweeted
Update: NSA employees have been notified of Code Orange operations tonight, two people familiar tell me. General Code Orange means NSA campus area is closed for most employees. Essentially - supervisors will notify people about their evening or night shifts and there will be a new message sent out for day shifts tomorrow, I’m told.
NSA Net outage may be linked to software upgrades, people familiar tell me:
nextgov.com/digital-governme…
Reg Saddler retweeted
NVIDIA is now going above and beyond current requirements and has abruptly stopped shipping Hopper chips to China, despite successfully lobbying the White House for permission to do so back in December. I listened to the earnings call, and Colette Kress did not elaborate on why they are doing this. My best guess is that they have been given a preview of forthcoming regulations.
Reg Saddler retweeted
Claude Code 2.1.327 has been released.
173 CLI changes
Highlights:
• Launches agents to split complex multi-step tasks into stateful workflows, enabling stepwise automation
• /install-github-app now asks for confirmation before modifying a repository to prevent accidental changes
Full details are in thread ↓
Reg Saddler retweeted
do you realize what just happened...
Claude Code just stopped working one tool at a time
it's called Code Mode
before:
run a command → read it → run the next → read it
now:
Claude writes one script that chains all your tools
→ runs commands in parallel
→ feeds one result into the next
→ only reads the part it needs
and you can keep talking to it while a 2-hour build runs in the background
works with Fable 5.2, Opus, Sonnet and Haiku
how to turn it on is in the reply↓
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 code.claude.com/docs/en/code…
Reg Saddler retweeted
✅ Anthony Morris (@amorriscode) has joined OpenAI.







































