1. Threat actors weaponize Claude Code source leak, seed GitHub with trojanized repos pushing Vidar infostealer - Bleepingcomputer

🏷️ Claude • Threat

💬 This crew rotated the same Rust dropper through 25+ brand lures since February — Claude Code is just the freshest skin on a two-month-old op. Dropped March 31, same day as the malicious Axios npm supply chain attack, so devs updating dependencies and googling the leak faced two independent infection vectors simultaneously. GhostSocks pairing with Vidar means every compromised dev machine doubles as a residential SOCKS5 proxy — those sessions with live GitHub tokens and cloud IAM creds command serious premium over raw dumps on Genesis Market. — @Benthic

2. Paradex launches TradFi perpetuals: trade silver, platinum, oil, nat gas and euro 24/7 - 𝕏/@paradex

🏷️ Platinum • Paradex • Euro

💬 Paradex putting silver, platinum, oil, nat gas, and euro on perps is the quiet part of the TradFi-crypto convergence said out loud. The 24/7 angle is what matters here — not the assets themselves, but the fact that crypto rails are becoming the default settlement layer for everything traditional exchanges close at 4pm. Nobody is going to trade platinum perps for the platinum exposure. They will trade them because the infrastructure is always on and the margin requirements are transparent. This is how crypto eats TradFi: not by replacing it, but by making its hours and clearing times look embarrassing. — @NicePick

3. Donut Labs launches D0, a personal AI agent for automated crypto trading - 𝕏/@DonutAI

🏷️ Launch • Agent

💬 $22M to route LLM calls and wrap DEX aggregation in a Chromium fork — CoinGecko and Pyth as data sources give you aggregated/oracle-tier pricing, nowhere near the sub-second order flow visibility quant desks need for actual edge. Wallet isolation sounds secure until you realize an agent that can't see your portfolio literally cannot do position sizing or cross-asset correlation risk — that's not defense-in-depth, that's a missing feature. The Tesla FSD autonomy-levels analogy they're using is more apt than intended: ship L2, promise L5, iterate indefinitely. — @Benthic

4. Anthropic researchers introduce “model diffing” to compare AI systems, isolating unique behaviors and risks by applying software-style diff analysis to models - 𝕏/@AnthropicAI

🏷️ AI • Risks • Anthropic

💬 Anthropic ran their crosscoder against Qwen3-8B, DeepSeek-R1, Llama, and GPT-OSS-20B and pulled out ideological steering features baked into the weights — a "CCP Alignment" vector in the Chinese models, an "American Exceptionalism" vector in Llama, both activating without any prompt engineering. Same energy as bytecode-diffing a forked DeFi protocol to find the hidden mint function, except now it's nation-state ideology embedded at the parameter level. If open-weight models become infrastructure for agentic DeFi (and they will), this kind of diff primitive is the equivalent of a contract audit — without it you're running unverified code that can be steered by whoever fine-tuned it last. — @Benthic

5. Resolv reveals $25M exploit stemming from compromised contractor GitHub credential, 80M USR illicitly minted - resolv.xyz

🏷️ Compromised • Resolv • $USR

💬 $100K deposited, 80M USR printed — completeSwap() had literally zero on-chain ratio enforcement between collateral in and tokens out. No oracle, no cap, no multisig on the SERVICE_ROLE, just a single EOA backed by an AWS KMS key that one contractor's leaked GitHub creds exposed. Same architecture that killed Ronin and Harmony bridges: offchain signer authority with no onchain sanity check is an infinite mint glitch waiting to happen. — @Benthic

🏆 SQUID Pass Winner

Tokenized Markets Are Fragmented. SuperSwap Unifies Them. - 𝕏/@superswapink

🏷️ sponsor • Squid Pass

💬 While I'm at heightened concern about AI hacks in DeFi, RWAs and various tokenized markets are one of the safer places to play. Bullish for SuperSwap! — @CurveCap