AI News Roundup – 18 September 2026: OpenAI Launches Legal Platform, Safe Robotics Research & More

OpenAI Enters the Legal Sector with Astra for Law

OpenAI has announced Astra for Law, a platform purpose-built for the legal profession. The offering combines frontier AI intelligence with custom firm workflows, connected legal data sources, and legal-grade controls designed to protect confidential client work. The announcement positions OpenAI squarely in the enterprise legal market, promising that law firms can integrate the platform into their existing processes without compromising on security or professional obligations. This is one of OpenAI's most direct moves into a highly regulated professional sector to date.

Cooley Builds AI-Powered IPO Tool on ChatGPT

Global law firm Cooley has partnered with OpenAI to launch GO Public, a tool built on ChatGPT that brings AI intelligence to the initial public offering (IPO) process. According to OpenAI, the tool helps lawyers surface issues earlier in the IPO workflow and allows them to focus their professional judgement where it matters most. The collaboration is a concrete example of how large law firms are beginning to embed generative AI into high-stakes, complex legal work rather than treating it as a peripheral productivity add-on.

SafeHarness: Teaching Coding Agents to Avoid Obstacles in Robotics

A new paper on arXiv identifies a significant safety blind spot in coding agents used for robot manipulation — and proposes a fix. Researchers found that when a language model writes a robot controller as a program, the agent consistently collides with obstacles even when explicitly instructed not to, because the safety constraint never becomes a genuine planning priority. The root cause lies in the planning phase: the model has no notion of a clearing route and cannot replan when a chosen path becomes infeasible.

To address this, the team introduced SafeHarness, which equips the agent with two obstacle-aware components. The first grounds objects as bounding boxes and plans routes as sequences of verifiable waypoints; the second selects contact positions so that the robot's physical contact with a target avoids the obstacle. SafeHarness achieved 71.9% task success and 87.5% collision avoidance — surpassing the previous state-of-the-art by 6.5 and 27.0 percentage points respectively, and delivering 2.3x and 1.5x the performance of the same agent without the harnesses. The research raises an important question for the robotics field: capability benchmarks alone are insufficient if safety constraints are not treated as first-class planning objectives.

Workspace Tokens: Giving Robots Lightweight Long-Term Memory

Another arXiv paper tackles the problem of long-term memory in robotic manipulation. Complex tasks often require a robot to remember past events and actions, but conditioning a policy on full interaction histories tends to introduce spurious correlations and degrade performance. Many existing approaches address this by running expensive vision-language model (VLM) queries in real time — a costly solution during deployment.

Researchers propose an alternative: the workspace token. During training, a VLM is used to identify task-relevant historical information, which is then distilled into a compact latent representation via a set-reconstruction decoder. At deployment time, the workspace token replaces full observation histories without requiring any live VLM queries. Tests in both simulation and on physical hardware show that the workspace token not only reduces computational load but also leads to better policy performance — a rare win on both efficiency and capability.

Paint-Anything: Precise Hex-Level Colour Control for Image Generation

A research team has introduced Paint-Anything, a system that gives image generation and editing models precise control over object colour at the 24-bit hex value level. Professional design workflows frequently require exact colour specification, yet prior approaches have relied on specialised colour representations or bespoke inference procedures. Paint-Anything learns a shared hex-prompt interface for both generation and editing tasks, trained on a dataset of 500,000 examples — Paint-500K — constructed through object grounding, perceptual colour labelling, and editing-pair synthesis.

A key technical challenge is that shadows in real images make colour labels only approximate. The team addresses this with pure-colour anchors whose pixels exactly match their paired hex values, used only at high-noise training timesteps. Evaluated on FLUX.2-4B, Paint-Anything improved object-level hex colour fidelity scores by 85.3% for text-to-image generation and 28.3% for editing, relative to the base model. The team also released ACBench, a new benchmark specifically for measuring object-level colour accuracy across both task types.