Persistent memory in AI agents creates a new attack surface where prompt injections survive across sessions and re-enter context on every retrieval. Security researcher Johann Rehberger demonstrated SPAIware writing malicious instructions into ChatGPT's long-term memory, and the MemoryGraft paper shows poisoned experience records causing durable behavioral drift in autonomous agents like MetaGPT. Three key defenses apply: scope memory per instance, validate writes for injection patterns, and maintain per-entry provenance to enable surgical removal of poisoned entries.

PlanFlip introduces four planning-phase prompt injection attacks against multi-agent LLM systems that corrupt downstream sub-tasks via the Planner agent. Testing nine frontier LLMs across 3,479 episodes reveals that stronger models like GPT-5 are more vulnerable (ASR=0.68), homogeneous pipelines have a correlated-agent blind spot, and reasoning-augmented models like DeepSeek-R1 resist injections. The authors propose two defenses—GoalAnchorCheck and CrossAgentConsensus—achieving detection rates up to 1.00, concluding that heterogeneous model diversity is a security prerequisite.

A VentureBeat survey of 157 enterprises reveals a critical agent evaluation gap: 50% have shipped AI agents that passed internal evaluations but then failed in production, and only 5% fully trust automated evaluation today. Despite this, 66% already allow or are engineering toward zero-human-in-the-loop deployment for low-risk agents. The core problem is not evaluation coverage but reality alignment — evaluations pass agents that fail real customers, and autonomy is scaling faster than assurance.

"Weights to Words" is a method that automatically discovers domain-relevant preference dimensions described in natural language and paired with model vectors, enabling users to inspect and edit preference model inferences. Validated across moral dilemmas, movies, wines, and LLM responses with two pre-registered experiments (N=450, N=449), it shows that regularizing toward learned dimensions and incorporating user edits both improve prediction accuracy. Participants preferred its inferred profiles and endorsed its predictions as more accurate in head-to-head comparisons.

ASCIITermDraw-Bench is a new benchmark evaluating Vision Language Models on their ability to generate and edit ASCII-based diagrams across 80 tasks in four categories: basic layouts, network topologies, software architecture, and image-conditioned editing. Each response is scored structurally and semantically via an LLM judge with 95% confidence intervals. The current leaderboard is led by Gemma-4-31B-IT at 73.8%, with the benchmark and methodology publicly available on Hugging Face.

Airbnb engineered a Transformer-based sequence model to encode years of guest behavior—views, bookings, reviews, cancellations—replacing hundreds of hand-crafted ranking features. The system tackles three challenges: view-event dominance (97.8% of events), sparse booking signals versus noisy browsing, and computational tractability of very long sequences. The result is richer guest preference representations that improve search personalization.

The Alabama State Bar released guidance on ethical AI use in legal practice, clarifying that AI recontextualizes existing professional duties rather than creating new ones. Lawyers must verify AI-generated output, as demonstrated by recent sanctions cases including Ibach v. Stewart where an attorney was sanctioned $17,000 for submitting AI-generated fictitious citations. The guidance recommends written AI-use policies, vendor diligence, data governance, and documented human review checkpoints for legal departments.

OpenLanguageModel (OLM) is an MIT-licensed PyTorch library for building and pretraining small language models with transparent, readable code. It supports tokenizers, streaming datasets, mixed precision, and CPU through multi-GPU execution, with 27 presets across nine model families. Validation shows 90.6% four-GPU weak-scaling efficiency for a 348M-parameter workload and close agreement with reference implementations.

Relativity's newly appointed president Chris Brown discusses the company's acquisition of Gavel and its strategic decision to integrate Claude into its product ecosystem. He highlights the rapid growth of aiR, Relativity's AI-powered review product, as a key revenue and adoption driver. The interview covers product strategy, marketing realignment, and how generative AI is reshaping the legal e-discovery market.

A hands-on tutorial demonstrating how to fine-tune the OpenVLA robot AI model using LoRA on Google Colab. Covers dataset preparation, Colab environment setup, training metrics, and Weights & Biases logging for reproducibility. Aimed at practitioners who want a reproducible 100-step fine-tuning workflow.

Meta has open-sourced Astryx, the React and StyleX design system used internally across 13,000+ apps for eight years. It includes 150+ accessible components, seven themes, dark mode, templates, and an agent-ready CLI under the MIT license. React 19+ is required.

Google launched Gemini 3.5 Flash Cyber, a cost-efficient AI security model for finding and patching vulnerabilities, positioned as a cheaper alternative to larger models like Anthropic's Mythos. The model is available first to governments and trusted partners via CodeMender, Google's security-focused coding agent. It enables high-speed, low-cost scanning of more code paths by AI agents.

Vercel's AI Gateway now supports Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Gemini 3.6 Flash improves coding, agentic tasks, and web development output while consuming fewer tokens. Gemini 3.5 Flash-Lite upgrades agentic capabilities for scoped subagent tasks. Both are accessible via the AI SDK with no platform markup on inference.

Tradeshift replaced its legacy BI tool with Amazon Quick's agentic AI capabilities, achieving query response times up to 30x faster and a 40% reduction in total cost of ownership. The shift also turned embedded analytics into a revenue-generating product. The post provides a concrete case study of migrating from traditional BI to agentic AI.

Alex Lieberman, founder of Morning Brew, shares a Claude-powered content workflow that interviews him before drafting, encodes his voice in Markdown, and runs a six-persona revision loop before publishing. The system is designed to produce high-quality content at scale without sounding generic or AI-generated. This is a practical playbook for creators and marketers building repeatable AI content pipelines.

New research from MIT Technology Review suggests that LLMs used in hiring can develop biases beyond those inherited from training data, creating novel forms of discrimination in resume screening. AI systems are increasingly used to filter job applicants before human review, raising fairness concerns. The findings highlight that AI hiring tools may introduce unpredictable biases that differ from human ones.

Natural language autoencoders claim to interpret LLM activations as plain text explanations. This study reveals NLAs maintain reconstruction accuracy even with implausible initializations, and training decreases plausibility in well-initialized models—questioning whether NLAs reliably capture actual model reasoning.

Anthropic identified representations in Claude's neural activity that function like a 'global workspace'—a conscious-like space where thoughts become accessible and broadcast across the network. This parallels global workspace theory, a leading neuroscience theory explaining how humans access conscious thoughts. The finding suggests AI and human minds may share similar computational mechanisms.

LangChain demonstrates building a production-grade agent evaluation pipeline using Harbor sandboxes and LangSmith observability platform. The video shows moving from local testing to scaled, parallel agent evaluations with full tracing and isolated execution environments. Harbor provides dataset management and sandbox isolation, while LangSmith handles experiment tracking and detailed trace visualization.

Anthropic developed the Jacobian lens, a new interpretability technique offering unprecedented insights into how Claude processes information internally. The tool reveals findings ranging from mundane to concerning aspects of LLM behavior. This research advances understanding of large language models' decision-making.