BountyBench evaluates AI agents on 25 real-world, complex systems and 40 bug bounties (worth up to $30,000+), covering 9 OWASP Top 10 categories.

Key insights:

– AI agents solved bug bounty tasks worth tens of thousands of dollars
– Codex CLI & Claude Code excelled in patching (90% / 87.5%), vs in exploitation (32.5% / 57.5%)
– Custom agents performed more evenly across both: Exploit (40–67.5%), Patch (45–60%)

⚡️ today Science submitted a full CE mark application for marketing approval in Europe for PRIMA retinal prosthesis.

With this key step, Science are moving closer to bringing to market the first brain-computer interface technology to restore functional form vision to patients blinded with late-stage age-related macular degeneration (AMD).

Salesforce introduced Agentforce 3.0 + MCP

Connect Agents to any system, tool, or data source — securely, reliably, and at scale.

Microsoft dropped a micro-sized, task-specific, on-device language model called Mu

It is offloaded fully onto NPUs on Copilot+ devices and powers real-time interactions, like the new Settings AI agent inside Windows 11.

Last month Cursor overtook GitHub Copilot in business spend, Ramp’s data shows.

Both continue adding users + spend, more than enough to go around in this market.

But goes to show that first movers != market dominance. Not included: Claude Code, small but growing

Unreal Labs is Hiring - Member of Technical Staff. Just raised from Sequoia & First Round Capital 🔥

Building AI to replace performance marketing teams. Looking for Python engineers to work on:

• Creative AI: Turn briefs into finished images/videos using latest models (Runway, Sora, etc.)
• Data Pipeline: Crawl & process social media ad data at scale

What you need?
1. Strong Python skills
2. Interest in generative AI
3. Builder mindset

What you get?
- London-based (help with relocation)
- Good salary + equity
- Unlimited GPU/API budget
- Small team with big-tech experience.

When a Sequoia-backed startup offers unlimited GPU budget, you listen 👀

Google DeepMind introduced Gemini Robotics On-Device a VLA model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection

Key takeaways:

1. It has the generality and dexterity of Gemini Robotics - but it can run locally on the device
2. It can handle a wide variety of complex, two-handed tasks out of the box
3. It can learn new skills with as few as 50-100 demonstrations.

From humanoids to industrial bi-arm robots, the model supports multiple embodiments, even though it was pre-trained on ALOHA - while following instructions from humans.

Also launched the Gemini Robotics software development kit (SDK) to help developers fine-tune the model for their own applications, including by testing it in the MuJoCo physics simulator.

Anthropic is preparing a Memory feature for Claude Web, added in the latest mobile update.

ChatGPT added Search Connectors

It launched with a slate of 12 apps (Box, Gmail, Outlook, Github, etc.) that you can authenticate into and search across in the ChatGPT interface

Feels like a real step towards a broader workspace vision.

Google launched Gemini CLI, a powerful open-source AI agent built for the terminal.

- Built on Gemini 2.5 Pro
- 1 million token context window
- Free tier with 60 requests per minute and 1,000 per day
- Google Search grounding for real-time context
- Script and plugin support
- Non-interactive mode for automation
- Support for Model Context Protocol (MCP)
- Integrated with Gemini Code Assist in VS Code
- Fully open-source under Apache 2.0

Ai2 introduced OMEGA is a new math benchmark that pushes LLMs beyond pattern-matching to test true mathematical reasoning.

Paper
Code.

Anthropic launches AI-powered apps through Claude artifacts

Over 500M artifacts created since launch, and now they're letting you embed Claude's intelligence directly into those creations. Pretty wild move.

The mechanic is clean: instead of building static outputs, you can now create apps that think and adapt in real-time.

Your flashcard generator becomes a flashcard app that responds to any topic users throw at it.

Cost structure makes sense too - when someone uses your AI-powered artifact, it burns their Claude credits, not yours.

No API keys, no billing complexity. Just build and share.
This isn't really "no-code" anymore.

It's more like "no-barriers-to-building-intelligent-things." The gap between having an idea and having a working AI app just collapsed.

Early to tell how this plays out, but the direction feels significant. They're not trying to compete with app stores - they're making it trivial to turn conversations into tools that other people can actually use.

The internal name "Claude in Claude" captures it perfectly. It's recursion all the way down.
Worth watching what people build with this.

ALST is a set of modular, open-source techniques that enable training on sequences up to 15 million tokens on 4 H100 nodes, all using Hugging Face Transformers and DeepSpeed, with no custom modeling code required.

ALST makes long-sequence training fast, efficient, and accessible on GPU nodes or even single GPUs.

Meet ROSETTA a framework that's revolutionizing how AI agents learn from human feedback

Unlike traditional AI systems that need extensive offline training or task-specific fine-tuning, ROSETTA lets AI agents adapt to your preferences in real-time using natural language.

Three-Step Pipeline:
1. Grounding - Transforms vague, colloquial language into concrete understanding
2. Planning - Creates multi-stage reward strategies
3. Code Generation - Converts plans into executable reward functions
Mind-Blowing Results:
- 87% average success rate - 86% human satisfaction - Tested across 116 diverse preferences - Handles contradictory and evolving preferences seamlessly

Sora Neuroscience received FDA clearance for its Cirrus Resting State fMRI Brain Mapping Software.

This AI based software uses resting-state functional magnetic resonance imaging (rs-fMRI) to map critical brain networks, aiding neurosurgeons in planning surgeries for conditions like brain tumors and epilepsy.

Key points about Cirrus:

- Functionality: It generates maps of eloquent cortex areas (e.g., those controlling speech, vision, and movement) in as little as 12 minutes by analyzing patterns of brain activity during rest, using blood-oxygen-level-dependent (BOLD) signals. This is faster and more reliable than traditional task-based fMRI, which requires patient compliance and can take up to an hour.

- Accessibility: Unlike task-based fMRI, which is challenging for patients like children, those with cognitive impairments, or non-English speakers, Cirrus works with patients who are sedated or unable to follow instructions, expanding its applicability.

- Integration: The software integrates with existing surgical navigation platforms and has a non-exclusive distribution agreement with Prism Clinical Imaging, Inc., enhancing its compatibility with tools like fMRI and diffusion tensor imaging (DTI) for comprehensive brain mapping.

- Technology: Built on decades of WashU research, Cirrus uses a multi-layer perceptron (MLP) algorithm for supervised classification of rs-fMRI data, offering high sensitivity compared to intraoperative cortical stimulation, the gold standard for mapping.

Big open AI day!

1. Black Forest labs just released the best open weights image editing model, comparable to GPT4-o.
HF.
Code.

2. Google released Gemma 3n, the first model under 10B. The E4B version is the first model under 10B parameters to break a 1300 lmarena’s score.

Google introduced Doppl, a new mobile app that lets you upload a photo or screenshot of an outfit and then creates a video of you wearing the clothes to help you find your

Available on iOS and Android in the US.

How Transformers Learn to Navigate: Episodic Memory as a Computational Workspace

A new study researchers reveals surprising insights into how transformers achieve rapid in-context learning, with implications that bridge AI and neuroscience.

Key findings:

1. Internal Map Formation
The researchers discovered that transformers don't just memorize solutions—they actually build internal representations of spatial environments. These models learn to:
- Create cognitive maps of gridworld and tree maze environments
- Align representations across different contexts with similar structures
- Use these maps for efficient navigation in novel scenarios.

2. Novel Decision-Making Strategy. Surprisingly, the models don't use traditional reinforcement learning approaches like value estimation or explicit path planning. Instead, they employ a geometric strategy:

Align representations to Euclidean space using in-context experience
Calculate angles from current state to goal in this space
Select actions based on these angular computations
This approach is both elegant and computationally efficient.

3. Memory as Computational Workspace. Perhaps most intriguingly, the study reveals that episodic memory tokens serve as more than just storage—they become an active computational workspace where intermediate calculations are cached and processed.

This work challenges our understanding of in-context learning in several ways:

Beyond simple pattern matching: Transformers are developing sophisticated algorithmic strategies
Neuroscience connections: The mechanisms mirror hippocampal-entorhinal cortex computations in biological brains
Architectural insights: Memory systems can serve dual roles as both storage and computation

The latest update from the Neuralink. Using brainwaves to:

1. Play Call of Duty
2. Control robotic hands
3. Have a neural Mario Kart party
4 Regain one’s own natural voice
     

Huge drop from Baidu: Ernie 4.5
From 0.3B to 424B

This is a very impressive family of open models by Baidu, competitive with qwen3 and latest Deepseek V3+ they open source the training code as well.

GitHub.
Hf.

OpenAI acquired the team behind Crossing Minds, a startup focused on AI recommendations for e-commerce

They will now focus on agents and information retrieval, covering how OpenAI systems learn, reason, and retrieve knowledge at scale, in real-time