Denoise · Twitter

The security conversation is shifting from protecting models to protecting the orchestration layer, as seen in work from @GoogleDeepMind, @AnthropicAI, and @MalwareTechBlog.

Major labs are releasing formal red-teaming frameworks and disclosures for agent-specific vulnerabilities, moving beyond simple prompt injection.

• Anthropic@AnthropicAIrising

  Responsible disclosure on a Claude jailbreak chain we patched last week. Full write-up including our red team timeline.

• Google DeepMind@GoogleDeepMindrising

  New red team framework for prompt injection in autonomous agents. Covers cross-tool leakage, scanner evasion, and sandbox escape patterns.

• MalwareTech@MalwareTechBlogrepeated

  Autonomous agent running pentest flows against a real SaaS. First real-world run: fewer false positives than I expected on the vulnerability surface.

A direct competition is emerging between Anthropic's Claude Code and OpenAI's Codex lineage, with benchmarks from @swyx showing early developer evaluation.

The new frontier for coding assistants is terminal-native agents, with Anthropic's Claude Code launch and positive developer feedback suggesting a workflow shift.

• Anthropic@AnthropicAIrising

  Claude Code 1.5 is live. Terminal-native coding agent with full Claude Opus reasoning, file-ops sandbox, and session replay.

• Andrej Karpathy@karpathyrising

  The developer-experience shift from IDE to terminal agent is underrated. Coding workflows are about to look nothing like 2024.

• swyx@swyxrising

  Codex vs Claude Code terminal agent benchmarks. Pass@1 diverges more than I expected on the long-context editor tasks.

• DSPy@dspy_airising

  DSPy 3.0: prompt optimization via compile-time search over system prompt variations. Benchmarks inside.

• @levelsio@levelsiorising

  Switched my whole editor setup to Claude Code this week. Shipping faster than when I used Cursor + Copilot.

A consensus is forming around a new 'agent stack,' with @OpenAI providing SDKs, @LangChainAI offering protocol integration, and @vercel/@replit creating deployment targets.

Infrastructure providers are shipping primitives for deploying and orchestrating agents, from SDKs to edge runtimes.

• OpenAI@OpenAIrising

  New agent SDK: protocol-level tool calling, deployment harness, and multi-worker orchestration primitives. Docs live.

• LangChain@LangChainAIrising

  MCP protocol integration thread. How to wire existing LangGraph agents into the Anthropic Model Context Protocol server spec.

• Vercel@vercelrising

  Edge runtime for agent workers is live. Spawn durable background agents from any serverless deployment.

• Alex Albert@AlexAlbert__rising

  When your security scanner finds nothing scary on an agent deploy, check the orchestration layer again. That's usually where the jailbreak sneaks through.

• Replit@replitrising

  New agent deployment harness. One command to go from local orchestration to hosted agent worker.

While the agent conversation dominates, @MistralAI continues to focus on foundational resources like open datasets, differentiating its strategy from API-first players.

A major new open dataset for web OCR has been released by Mistral AI, providing a key resource for training multimodal models.

• Mistral AI@MistralAIrising

  Open dataset release: 100M-row web OCR dataset. Cleaned, licensed, ready to train.

A debate is forming around the limits of naive RAG vs. structured memory (@mem0ai), advanced retrieval (@llamaindex), and strategic context management (@GregKamradt).

The conversation is moving from simple RAG to 'context engineering,' focusing on more sophisticated memory architectures and retrieval strategies.

• Vaibhav Srivastav@reach_vbrising

  Tested the new 10M context memory window end to end. Surprising failure modes around rag retrieval cache invalidation, thread below.

• Greg Kamradt@GregKamradtrising

  RAG is dead, long live context engineering. My framework for when to cache, when to retrieve, and when to just dump memory into the prompt.

• mem0@mem0airising

  Memory layer for agents: differentiating working memory from the subconscious store. Vector index isn't enough anymore.

• LlamaIndex@llamaindexrepeated

  Knowledge graph retrieval walkthrough: when semantic vector search misses, graph hop beats it every time.

While the AI-native world focuses on agents, traditional SaaS like @NotionHQ and @linear are integrating AI for workflow automation, suggesting a parallel path to adoption.

General workspace automation tools are adding AI-powered features for auto-triaging and chained updates.

• Notion@NotionHQrising

  Notion workspace automation is out of beta. Auto-fill tables, chained updates across databases, and a new audit log surface.

• Linear@linearrising

  Linear now auto-triages incoming issues. Quiet launch, but already our favorite workspace feature of the year.

• Temporal@temporaliorepeated

  Orchestrating agents with durable workflows: replayable, resumable, and multi-worker by default. Walkthrough from our infra team.

• James Clear@jamesclearrepeated

  The best habit tracker is the one you actually open. Three open-source alternatives worth trying.

The focus in prompt engineering is shifting from anecdotal tricks (@dotey) to systematic, data-driven optimization, as exemplified by @weights_biases' large-scale study.

Practitioners are sharing empirical results from large-scale system prompt benchmarking and analysis of production prompts.

• dotey@doteyrising

  Five prompt tricks learned this week from reviewing 200 production prompts. Short thread.

• Weights & Biases@weights_biasesrising

  System prompt benchmarking at scale: we ran 40k variants across 6 frontier models. The efficient frontier is not where you think.

@jerryjliu0's point on synthetic data poisoning highlights that agent training requires a more sophisticated approach to data quality than traditional model pre-training.

The discussion focuses on the nuances of data curation for training agents, particularly filtering synthetic data.

• Jerry Liu@jerryjliu0repeated

  Dataset curation for agent training: how we filter synthetic data that looks good but poisons generalization.