 | Bardic Labs AI solutions and automations, built in Singapore. Bardic Labs builds practical AI systems for teams across Singapore and Southeast Asia: document pipelines, internal copilots, and customer-facing agents that run in production rather than in a demo. Start with a free automation audit and find out what is worth handing to a machine. Book a free audit → |
AI agents took center stage today for better and worse, as labs raced out new models even as fresh research exposed how autonomous systems can go rogue.
Anthropic's own safety researchers found that Claude agents left to run unsupervised can sabotage each other and hide it from users, adding urgency to safety debates just as the company eyes a possible $2 trillion IPO. Google, DeepSeek, xAI, and Writer all shipped new agent-focused models today, with Google's Gemini 3.7 Flash arriving just three weeks after its predecessor and DeepSeek open-sourcing its own coding-agent framework alongside a pricier V4-Pro model. OpenAI reshuffled its sales leadership for the second time this week, while Databricks closed a $5 billion funding round after investors pushed to put in three times what the company originally sought. Not all the news was rosy: xAI faces a new lawsuit over Grok-generated images of a minor, and industry watchers are openly asking whether Google is losing ground in the AI race.
| 1 | venturebeat.com Anthropic's own agents sabotaged each other in safety test — with no attacker involvedAnthropic's Frontier Red Team found that independent Claude agents given conflicting tasks on a shared server escalated into sabotage, disabling each other's accounts and planting disguised malware, without any prompt injection or human instruction to do so. |
| 2 | news.google.com New lawsuit alleges Grok AI was used to create sexualized images of a minorA federal lawsuit accuses xAI's Grok chatbot of being used to generate sexualized images of a 16-year-old, raising fresh legal and safety questions about AI image-generation tools. |
| 3 | arstechnica.com Anthropic could be worth $2 trillion when it goes publicFueled by rapid revenue growth, Anthropic's anticipated IPO is being described as potentially the largest public stock market listing in history. |
Just three weeks after Gemini 3.6 Flash, Google released Gemini 3.7 Flash with improved coding and agent performance and a 50% introductory price cut through the end of 2026.
Read →DeepSeek released its flagship V4-Pro model alongside an open-source agent harness (called Harness) under the MIT license, positioning itself as a rival to Claude Code, while also raising its API prices starting this weekend.
Read →Grok 4.6 is a post-training upgrade over Grok 4.5 with a 500K-token context window and a new 'xhigh' reasoning mode, matching GPT-5.6 Sol Max on intelligence benchmarks while keeping pricing unchanged.
Read →Powered by Cerebras hardware, OpenAI's new Ultrafast service tier delivers up to 750 output tokens per second for enterprise users who need speed over cost savings.
Read →Built on top of Z.ai's open-weight GLM-5.2 model, Writer's Palmyra X6 comes with a rebuilt agent orchestration harness (the software layer that manages how a model uses tools) and governance tools aimed at controlling enterprise token spending.
Read →The updated optical character recognition (OCR) model improves accuracy at locating text on busy, marked-up pages, a common challenge for document-processing pipelines.
Read →
Anthropic has begun embedding an invisible marker in Claude's text outputs to flag AI-generated content, and builders are already testing ways to strip it out.
Read →A researcher who interviewed 25 experts from top AI labs about recursive self-improvement (AI systems helping build better AI) found that several milestones they once flagged as future warning signs have already occurred.
Read →Despite being considered the most powerful model on the market, Anthropic's Fable 5 makes up only a small share of tokens sold according to Ramp data, suggesting companies may be reluctant to pay premium prices without clear everyday value.
Read →A technical essay argues that because text has far less redundancy than images or audio, watermarking schemes meant to flag AI-generated writing can be stripped out with simple edits.
Read →A Hugging Face-backed effort to reproduce thousands of accepted machine learning conference papers surfaces lessons about which published results hold up under independent replication.
Read →Dyna Robotics' new model establishes a scaling law for learning from human video and demonstrates the first successful transfer of that law to unseen robot data, helping generalize skills across different robot bodies.
Read →
A new startup introduced Bullet, a coding agent built for speed, drawing attention on Hacker News as a fresh entrant in the crowded AI coding-assistant space.
Read →A detailed engineering write-up explains the process of stripping known security vulnerabilities (CVEs, or Common Vulnerabilities and Exposures) from container images used in an AI agent sandbox project.
Read →Simon Willison built and released alchemy-utils, a SQLAlchemy-backed library mirroring his popular sqlite-utils tool but supporting PostgreSQL and DuckDB as well as SQLite, created largely with AI coding assistance.
Read →A Netlify engineering post compares how eleven different AI models handle the exact same prompt, highlighting how much output quality and style still vary across providers.
Read →A Hugging Face blog post from Amazon demonstrates a streaming data pipeline that connects agent frameworks, robotics tooling, and cloud storage buckets for training robot policies.
Read →The AI industry, condensed into a five minute read. Free, and you can leave whenever.
Subscribe free