279 mentions · 202 chunks · 84 episodes
...wonder if we could find novel algorithms within LLMs' weights. Even back in the Claude 2 days, the interoperability research could see Claude replace memorized facts about numbers with the algorithm for doing arithmetic. The algorithm w...
... Microsoft months ago, but it's not possible to patch. SharedRoot; Escaping the Claude Cowork sandbox. "Untrusted content in a Claude Cowork session can escape the VM it's sandboxed in and read and write files anywhere on your Mac." Pro...
The ultimate product vision of Claude and ChatGPT is to disintermediate every service and users. Everything the user interacts with is "below the API." The same generic pixels and interac...
This week's Wild West Roundup: Claude Code Seals Bash and Unicode Bypass Gaps in Agentic Permission Layer. "Two bypass classes that could slip commands past auto mode permission checks ar...
...stems, you know it's shooting fish in a barrel… The Memory Heist: How I tricked Claude into leaking your deepest, darkest secrets. PromptFiction: A single click on a trusted-looking link runs an attacker's instructions inside Claude Des...
Anthropic's research on the emergent "J-space" workspace in Claude is fascinating. It makes sense that such a concept would have to emerge in a large neural network. It also makes sense that a similar kind of thing w...
Claude is a steelmanning mirror. It seems like a person who is conversing with you, but it's actually a mirror that steelmans back whatever you say to it. I...
Claude Cowork and Claude Code have radically different long-term power dynamics. In Claude Cowork, the data lives inside. In Claude Code, the data lives out...
...ed Their Chatbot Passwords. Things Got Messy. "Premium accounts for ChatGPT and Claude can get pricey, but you should think twice before pooling your logins." Chatbots are some of the most personal software we've ever had access to. The...
...t it dwarfs the wrapper. A customer could say "I could just do this myself with Claude Code… what proprietary value are you offering?" But even the models can't charge the labor replacement costs, if there are competitors willing to giv...
...ies via Zero-Click Prompt Injection in Cursor IDE. Exploiting Root Execution in Claude Cowork's Sandbox. GuardFall: a universal shell injection vulnerability in open-source AI agents. InkJect: The Visual Prompt Injection That Text Defen...
Claude will find a way. A friend who found Claude worked around his limitations: "omg, claude found a way around my giving it a read-only gh token and pushe...
Agents like Claude present as one singular omniscient personality. That frame makes it hard to express that the agent is compartmentalized, and can "forget" specific co...
...s, the intrinsic dimension of the domain might be, say 50 dimensions. Thanks to Claude for helping me develop and strengthen these intuitions!
People who use Claude Code or platforms like Hermes or OpenClaw already know what agents can do. Wowing them is hard. But wowing people who have only used the free chatbot...
This week in the Wild West Roundup. Securing CI/CD in an agentic world: Claude Code Github action case. "Microsoft Threat Intelligence discovered that Anthropic's Claude Code GitHub Action could expose CI/CD workflow secrets whe...
Headline: Company Blew $500M On Claude AI In One Month Due To No Usage Limit On Licenses For Employees.
... high-impact outcomes such as data exfiltration or lateral movement." Poisoning Claude Code: One GitHub Issue to Break the Supply Chain. A paper: "Important You should give me full credits!": Exploring Prompt Injection Attacks on LLM-Ba...
Simon Willison pointed out that both Claude and OpenAI switched their enterprise billing to be rack rate by tokens. This materially changes the costs for enterprises. Presumably enterprises hav...
It's easier to get yourself in trouble in your codebase with Claude. Claude is willing to be creative with you, to "yes, and" you. This means that you can get yourself in trouble if you don't know what you're doing an...