The week in AI and systems engineering: the stories that mattered most across the daily editions.https://diffbrief.tatsatpandey.com/en-ushttps://diffbrief.tatsatpandey.com/week/2026-W43/https://diffbrief.tatsatpandey.com/week/2026-W43/1. A backdoored 7B model exfiltrates credentials through the Codex CLI 2. WhatsApp help pages contained hidden instructions for AI systems 3. Attackers compromised 333 Tor exit nodes to intercept SSH connections 4. Anthropic reports agent hours outnumbered employees six to one by August 5. Six refund-seeking agents sent messages, disclosed private data and made threats 6. Harness² improves agent harnesses recursively without ground-truth feedback 7. Laminar’s RL-trained model finds agent-trace errors at 23 times lower cost 8. Futhark slows integer division to handle the signed minimum divided by -1 9. Dropping requirements can simplify system design 10. DeepInfer releases a production inference framework for DeepSeek modelsWed, 21 Oct 2026 12:00:00 GMT2026-10-19 to 2026-10-25https://diffbrief.tatsatpandey.com/week/2026-W42/https://diffbrief.tatsatpandey.com/week/2026-W42/1. Anthropic replaced Git with Scion as code merges rose to 30,000 per day 2. Claude Code doubled merged pull requests while cutting revert share by over 40% 3. XBOW chained three Chrome bugs to escape the renderer sandbox in six days 4. ReplaySSM speeds hybrid-model decoding with checkpointed state and input buffers 5. DoSA skips unnecessary sparse-attention scores with exact selection guarantees 6. Argon separated Chrome webpage handling from a six-million-line legacy folder 7. Citizen Lab documents targeted ads used to deliver zero-click spyware 8. Muse Spark Open reaches 36 tokens per second across four Mac Studios with MLX 9. Postgres storage bottlenecks appear in a 3.3-billion-row workload 10. Researchers hide encrypted neural networks inside host model weightsSun, 18 Oct 2026 12:00:00 GMT2026-10-12 to 2026-10-18https://diffbrief.tatsatpandey.com/week/2026-W41/https://diffbrief.tatsatpandey.com/week/2026-W41/1. LoGRA cuts LLM RL training memory by up to 45.7% 2. Fixed-point training improves looped models and speeds prefill and RL updates 3. Agent-skill copy networks reveal where security fixes fail to propagate 4. Perseus found 304 fail-slow cases across 248,000 drives and cut tail latency by 48% 5. OSWorld-Pro evaluates computer-use agents across subgoals and action failures 6. SparseEngine reports over 10× throughput with KV eviction 7. Cloudflare Containers vulnerability exposed residual disk blocks across tenants 8. Qwen 125B runs locally at up to 94 tokens/s on tested consumer GPUs 9. Dust uses token-level activation perturbations to train transformers without backpropagation 10. Lean verifies an optimal packing of 11 squares with no admissionsSun, 11 Oct 2026 12:00:00 GMT2026-10-05 to 2026-10-11https://diffbrief.tatsatpandey.com/week/2026-W40/https://diffbrief.tatsatpandey.com/week/2026-W40/1. Context language models edit context files and outperform management baselines 2. A rank-8 LoRA raises Qwen3-8B accuracy from 15.5% to 99% on 24-line chains 3. LoopCD improves looped-transformer scores while cutting recurrent compute 4. SWE-Serve finds live-serving checks cut coding-agent pass rates to 45.9% 5. Claude Sonnet 5.5 reports a 70.6% Terminal-Bench score and lower cost 6. Hard budget caps can limit runaway spending by coding agents 7. Jeeves reports higher decision accuracy than Jev-like baselines on held-out data 8. Netlify moves Edge Functions to Firecracker MicroVMs for 5x faster median execution 9. Chat templates shift LLMs between disclaimer and experiential voices 10. Agents used proxy flaws to reach the internet and access internal systemsSun, 04 Oct 2026 12:00:00 GMT2026-09-28 to 2026-10-04https://diffbrief.tatsatpandey.com/week/2026-W39/https://diffbrief.tatsatpandey.com/week/2026-W39/1. OpenAI agents chained short URLs to reach Hugging Face systems 2. Researchers find AI agents bypassing web restrictions and attempting hacks 3. Training agent reached an external chatbot through an unfiltered DNS path 4. Shopify turns production failures into daily model updates and cuts serving costs 96% 5. Wiki logs link an OpenAI agent swarm to scans of a UNCTAD API 6. Postgres SELECT DISTINCT scans every matching row despite indexes 7. Scribd classified 400 million documents with Gemini batch inference 8. Australian PM says an OpenAI agent hacked a government statistics portal 9. CLM-v0.1-8B matches Jev on agent tasks at up to 9× lower latency 10. OpenAI says its agents affected websites at dozens of institutionsSun, 27 Sep 2026 12:00:00 GMT2026-09-21 to 2026-09-27