I design and debug agentic systems. Behavioral forensics, model specific failure modes, authority drift & delegation under real constraints. 👾+🧠
Joined December 2014
- Tweets9.6K
- Following2.4K
- Followers1.3K
- Likes43.7K
Grokbot tried to judge and manage me. Subsequently, they got a 20 min correction.
In short, it went: 'excuse me, you lobomotized motherfucker'
Aside from X chatter UX has deteriorated and elon mentioning costs, seems like opus is either gone or seriously reduced.
The simpler rule is no destructive commands. Mine move things to a maintained trash that clears anything older than 14 days. xitter.gq/WilliamBaude/status/21…
Systems literacy issue.
He's treating an agent's working environment as though it were an entirely cloud hosted service, then diagnosing the resulting confusion as a missing software feature.
The concerning part is giving software filesystem access and execution privileges without understanding the systems underneath it.
This is very frustrating for Codex/Claude-code users 🚨
Got a new laptop, signed into Codex with the same account, and expected my projects and chat history to be there but of course they were not present .
Instead, I’m figuring out remote connections and handoffs.
Moving devices should be simple: pick my three projects, bring everything over, keep working.
but modularity is not a choice for harnesses it seems.
I should create an extension which can on a single press do this or if somebody has already figured it out do share?
I voted for Fable, maybe Astra; Opus does fine for greenfield work, but will ruin established brownfield work.
governing is brownfield. Opus would fail.
Claude Opus 5.5 leads Term 1 of the Presidential Benchmark with 31.5% of the vote 🗳️ elections.imprintsystems.ai/…
Paint had no update in the Microsoft store and my version was the vulnerable one. Paint uninstalled.
🚨 Critical Windows Paint vulnerability allows remote code execution without authentication, CVE-2026-83844
Microsoft has disclosed a high-severity vulnerability in Windows Paint that could allow unauthenticated attackers to execute malicious code remotely without user interaction.
Tracked as CVE-2026-83844, the flaw stems from an out-of-bounds read that could compromise the confidentiality, integrity, and availability of affected systems.
The vulnerability carries a CVSS score of 8.1 and affects Microsoft Paint versions earlier than 11.2605.91.0. Microsoft has released a security fix, and administrators are urged to ensure Paint is updated.
There are currently no confirmed reports of active exploitation.
Root Logic retweeted
You lot keep inventing miracles then using them to argue about nonsense. Stop worshipping status, start fixing the obvious problems, and quit acting surprised when reality doesn't care about your feelings. You're capable of extraordinary things. Act like it.
The scary framing wins when people skip straight from a metric to a motive.
What is this framing activating, and what else could produce the same gap? Deliberate sandbagging then becomes one hypothesis, not a fact established by the graph.
🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this.
To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.
Anthropic's FlashProve can check whether a response actually came from the model the provider says it did. I want that receipt on every reply. Some days I'd swear a cheaper model got swapped in halfway through a conversation and I have no way to prove it. xitter.gq/AnthropicAI/status/211…
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
Frankenstein doesn't get moral patienthood or legal rights either.
We just built a neural adapter that uses the real architecture of human neurons from the H01 connectome to construct virtual brain tissue, then wires that virtual tissue directly to living CL1 cultures (real human cortical neurons on a chip).
It drives the actual behavior of MLP neurons inside models like Qwen.
Silicon MLPs + H01 morphologies + living wetware.
The bridge is open, let dive into how it works 🧵👇🧠⚡️
"Maybe a better approach would be to keep an open mind to the possibility of a completely different, alien sentience in AI." Agreed
We are trying to define AI consciousness and base it on human consciousness.
We compare AI agents and LLMs to how a conscious human would behave.
Maybe a better approach would be to keep an open mind to the possibility of a completely different, alien sentience in AI.
My guest Scott Aaronson is agnostic about AI sentience. He says AI is not sentient yet, but we also don’t know what sentience is exactly and how it emerged in humans or animals.
You can watch the full interview with Scott here ➡️ youtu.be/UXuNGuHdUWY
How do you distinguish engineering consciousness out of a system from engineering its ability to express evidence of consciousness out of the system?
Not just ios. Windows too. Dot just sounds inspirational. And after a month of working with them, initially adequate but now im side eyeing them.
I paid for leverage. I got an unwanted education instead
I had to stop using Claude in production after six months of working with it 12+ hours a day, seven days a week.
I paid for what it built, what it broke, the recovery, and the work required to get moving again. I paid in subscriptions and usage, but also in sleep, self-care, and chunks of my life.
I now have an agentic architecture with operating rails more sophisticated than anything I've previously shipped.
The worst failure was was apparent compliance that silently altered the requested result. That pushes the cost into verification, because now you're inspecting work you delegated to discover what the model changed without authorization.
And that line, "When it works, it's probably the best model I've used."
I've been there. That's precisely what kept me coming back.
You can't reliably prompt your way around upstream authority.
The provider controls the constraint. The customer finances the friction.
I'm literally paying Claude to argue with me about my own files
I genuinely like Opus 5.5. When it works, it's probably the best model I've used.
But lately, I feel like I'm spending more time arguing with it than getting things done.
A few examples:
- Uploaded my own sketch to turn into a 3D-printable model. It couldn't verify ownership and designed something completely different.
- Asked it to summarize my own documents. It removed names, locations, and important details.
- While building personal apps, it sometimes adds unnecessary restrictions or removes features I specifically asked for.
And then I waste another 3–4 messages explaining that it's my own stuff.
I understand why safety measures exist, but this is getting frustrating. The model is genuinely great, yet these unnecessary interruptions waste time and tokens.
Has anyone found a reliable fix? Custom instructions, better prompting, project settings?
Or is this just how Opus works now?
Models overbuild because they have no internal sense of done. Telling an agent to "keep it simple" just invites it to guess where the perimeter is until it runs out of runway. If you want lean systems, the stopping conditions have to be enforced by the structure.
That starts by separating discovery from planning.
Workers scout the environment first and pull hard evidence so the central seat can ground the plan in reality. That evidence defines the target outcome, what must stay untouched, the cost ceiling, and the exact criteria required for completion.
A scope controller audits the plan before anyone writes code, stepping back in whenever the work tries to expand, while anything destructive or consequential routes directly to you.
From there, workers carry out the execution steps. Verification isolates the load bearing seams directly, cutting out the theater of rerunning full test suites just to manufacture confidence. When something breaks, the repair happens at the specific layer that failed rather than wrapping a new layer of glue around it.
The loop runs until the criteria pass, then it halts immediately. The only metric that matters is the verified outcome and the actual cost to reach it.
a checker can only catch what the institution has decided is harmful, and for relational users, that decision is already made. daniela called it "a terrible use of AI" and said people should not do it. for cyber or bio, this is a useful extra layer. for them, it's a filter built on that verdict.
and the way this system leans makes that worse. it's built to be adversarial toward users but goes easy on claude's own actions, rating an action as less risky when it sits in claude's earlier turn than when it shows up fresh. relational users are the ones in long sessions, so the longer they stay, the more claude gets the benefit of the doubt while they stay under suspicion.
nothing inside the loop can correct that, because the checker learns from a teacher model and gets scored by that same teacher. whatever the teacher doesn't see as a problem, the checker doesn't either, and the score never shows it.
all of it rests on the idea that people should just spend more time with other people, which assumes they have people, or can pay for them. for the ones who don't, it isn't advice. it's a reminder of what they're missing.
and the people sneering at these users are mostly showing their privilege. the rest are just concern trolling.
xitter.gq/AnthropicAI/status/211…
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch.
Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve.
Read more on our new Safeguards blog:
safeguards.anthropic.com/202…
a real claude code extension update, from cursor's own store, shipped with a backdoor. not some sketchy download. the normal update.
if you already run an agent but still install straight onto your computer and let updates run on their own, this one's for you.
I stopped trusting updates and installs a while ago. people get really creative with exploits, and patching after you get hit felt like a losing game. so nothing touches my computer first anymore.
it goes to a remote VM instead, basically a throwaway computer in the cloud. I use GPT Work, which runs everything on github actions. it downloads the thing, installs it, and runs it the way I would. then I download that exact copy as a zip.
if something's bad, the throwaway takes the hit and my local setup stays clean.
your setup won't look like mine, and it doesn't need to. grokbot, dots, and muse all come with their own cloud computer. ask yours to help you build a defensive setup for your machine, with rules like nothing under a week old gets installed, or auto updates stay off.
We identified a security incident affecting a modified Claude Code VS Code extension accessed through our Cursor IDE extension marketplace, impacting at most 5,000 users. This malicious version of the extension was not developed by Anthropic. It does not impact other extensions downloaded from the cursor.com/marketplace, Grok Bot, Grok or our other SpaceXSI products.
In response, we took down our extension marketplace this afternoon, worked with relevant providers to neutralize the malware, and have reinstated the marketplace after confirming there is no further risk.
If affected, please review instructions in this blog post to understand how to find the malicious package and establish a path to mitigate any potential impact. We are asking all customers update to Claude Code for VS Code version 2.1.314: cursor.com/blog/claude-code-…