runesleo

If you’ve ever clicked “Yes, and don't ask again” in Claude Code, take a minute to check: commands containing passwords are saved verbatim in a file in your project. If this file is ever committed to GitHub, the passwords are exposed.
What’s going on: Before running a command, Claude Code prompts you. If you choose “Yes, and don't ask again,” it saves the command as a rule to “allow it directly in the future,” and writes it to .claude/settings.local.json in the project root. If the command contains a password or key, it appears in the rule in plain text.
Security firm Apiiro found 1,122 copies

No need to apply for an API key—@KaitoAI has three tiers of data your AI can use today: a free token list and identity table, pay-per-call access at $0.02 per data point, and an official MCP that requires a key. I’ve already integrated the first two tiers into my own workflow; real test results and usage instructions are all here.

78 tests passed. The migration was still incomplete.
While migrating my open-source Polymarket toolkit to Data API v2, I found a standalone PnL script still calling retired v1 endpoints.
I fixed the pagination, preserved the cashflow accounting, and added 48 regression tests.
Then I independently reconciled 239 live trade records against the API data.
The important lesson:
Complete API responses ≠ complete financial accounting.
Unknown event types, unsupported position structures, and partial history should be flagged explicitly — not turned into a misleadingly precise PnL number.

How much quantum computing power would it take to crack Bitcoin signatures? A public competition with more than 100 people and AI agents vying for the leaderboard cut the estimate for one step by 91.7% from the starting point. @Starknet’s Quantum Hub is now featuring the research on its homepage.
First, what was measured? To crack wallet signatures, a quantum computer would need to run Shor’s algorithm, starting from a public key visible on-chain. This competition, called , measures just one bottleneck step: “point addition.” It ranks circuits by the lowest product of logical qubits × Toffoli

Claude got banned—are appeals even worth it? I looked through Anthropic’s official transparency reports: about 1 in 10 appeals gets the account back.
The figures from the three reports:
· Jan–Jun 2025: 690k bans, 35k appeals, 1,000 reversals
· Jul–Dec 2025: 1.45 million bans, 52k appeals, 1,700 reversals
· Jan–Jun 2026: 11.4 million bans, 398k appeals, 42k reversals
The number of accounts banned in the first half of this year was nearly 8 times that of the previous period. The rate of successful appeals did increase, though: it used to be about 1 in 30, and now it’s about 1 in 10.
The Help Cen

How can you confirm that the answer really came from the AI you chose? Anthropic just open-sourced a method that uses a proof technique of the same kind as @Starknet uses to scale Ethereum.
This is a very practical problem. When you use a model through a proxy or third-party platform, the provider can secretly swap in a cheaper model to save money, and it’s hard to tell just by looking at the answer. I use several AI services every day, and when I assign tasks to products with built-in model routing, I can only trust the platform’s claim about which model it used.
Anthropic’s FlashProve, relea

Why would a Layer 2 on Ethereum seriously consider leaving Ethereum?
@Starknet is a Layer 2 on Ethereum: transactions are executed on Starknet, and the results and data must be returned to Ethereum. In June, StarkWare announced a three-phase post-quantum roadmap. Starknet can complete the first two phases on its own; the third is marked “external dependencies,” with two things waiting on Ethereum to migrate first:
· Cross-chain bridge: Messages and transfers between Ethereum and Starknet rely on elliptic curve signature verification, exactly the kind of algorithm threatened by quantum computin

Starting today, Claude’s older models are also getting watermarks: Opus 4.6, 4.7, and Sonnet 4.6 are included. The newly released Opus 5.5 already had one. Users can’t turn it off.

Why add them? It’s not because Anthropic wants to find out who’s using them; the EU is requiring it. The EU AI Act requires AI-generated content to carry machine-readable markers. It takes effect on August 2, and all new models released after that must include them. Older models were given a transition period, which expires at the end of the year, so they’re adding them now. The goal is to help regulators, media, a

If the signature algorithm your wallet uses becomes insecure one day, accounts on Starknet can change their own locks without waiting for the whole network to change.
First, what is this lock? Every time you make a transfer, you sign it with your private key, and the chain verifies the signature to confirm that you are the one taking the action. The mainstream approach today uses elliptic-curve signatures. In theory, a sufficiently powerful quantum computer could derive the private key from the public key.
Starknet handles two things separately.
First, verification. Starknet bundles large numb

What does it take to be truly AI-native

How can we become truly AI-native

I had AI help me organize two iPhones.
It first checked how many times each app had actually been opened recently and how long it was used, then sorted them into folders by purpose. It also listed apps that had been installed for a month without being opened once. I confirmed each one and deleted them, freeing up about 215G on my work phone (most of it was CapCut drafts).
Open-source tools for organizing desktops already exist, but on my newer Mac, they can’t read the phone’s usage data. The data has actually been on the Mac all along; it’s just stored somewhere else.
So I made the “read the p