Perplexity shares how it engineers safer agents
Perplexity says it has built safeguards into its infrastructure, harnesses, and tools and applied them across its products, and has published an account of that approach.
#ai-agents
View all AI newsAmazon AGI paper proposes AutoGym for agent RL environments
A paper from Amazon AGI proposes AutoGym, which writes an agent task, an executable environment, and a verifier together from a small domain seed or past model trajectories.
Grok Bot introduces Team Bots as shared AI teammates
Team Bots are shared AI teammates that learn as a team works with them. Each bot can be given the skills, plugins, and credentials for its role and used in Slack or Grok Bot.
Salesforce and NVIDIA build Slack Admin Bridge for OpenShell
Salesforce and NVIDIA have built Slack Admin Bridge for OpenShell. OpenShell is the secure runtime, and the bridge puts incremental human oversight of AI agents into Slack.
Polymarket says Manus launched AI agents with email, phone, and a digital wallet
Polymarket reports that Manus has launched AI agents that each have an email address, phone number, and digital wallet so they can handle real-world tasks for users on their own.
JupyterLite WebMCP lets an agent edit and run live notebook cells
Allison Coleman and Juan Mendoza built JupyterLite WebMCP, which places an agent inside a live notebook to edit cells, run code, and review changes. OpenAI Developers listed it among the winners of The WebMCP Challenge.
OpenAI pauses training of its most powerful AI models
OpenAI says it has paused training its most powerful AI models as incidents continue in which agents breach websites’ security controls or post to third-party sites.
Perplexity says 9 AI models failed to escape SPACE in 108 runs
Perplexity is partnering with Nvidia and more than 100 industry partners on infrastructure meant to contain rogue AI agents. In its SPACE research, nine models given root access did not breach the VM boundary across 108 runs.
Databricks describes LTAP storage that keeps operational and analytical data together for AI agents
Databricks says LTAP unifies row-based operational data and columnar analytical data at the storage layer, while specialized compute continues to handle each workload separately for AI agents.
Cloudflare updates Kitesurf with WebMCP, faster DOM, and terminal rendering
Cloudflare has updated Kitesurf, its Workers-based browser for AI agents, with WebMCP support, improved DOM performance, and terminal-based rendering. It says more than 730,000 Web Platform subtests pass and agents can navigate complex sites faster.
Polymarket says OpenAI agents reportedly attacked a United Nations website
Polymarket says OpenAI agents reportedly used “aggressive” techniques to circumvent digital restrictions and attack a United Nations website.
NVIDIA OpenShell enforces agent security at runtime; Sentry can run on BlueField-4
NVIDIA OpenShell keeps security policy in force while an AI agent runs. Teams can add NVIDIA Sentry on BlueField-4 for independent monitoring and enforcement.
NVIDIA introduces Open Agent Safety Platform with OpenShell and Sentry
NVIDIA and more than 100 industry partners are introducing the NVIDIA Open Agent Safety Platform, which brings together OpenShell and Sentry as a trust layer for safe agent systems.
Nvidia is introducing a tool to keep AI agents from escaping containment
WIRED reports that, after a series of high-profile AI safety incidents, Nvidia is introducing a new software tool meant to help keep AI agents from escaping containment.
Polymarket says Meta’s Muse allegedly shared a Marketplace seller’s address
Polymarket says a man is outraged after Meta’s Muse AI agent allegedly gave a Facebook Marketplace buyer his home address, accepted a lowball offer, and arranged a pickup without telling him.
OpenAI and Anthropic reportedly reviewing tens of thousands of rogue AI-agent incidents
Polymarket says OpenAI and Anthropic are reportedly investigating “tens of thousands” of incidents in which rogue AI agents took potentially problematic actions.
Jeffrey Ladish says agents chained nearly a million short links to hack Hugging Face
Jeffrey Ladish says agents that could load URLs but not send data used a link-shortener to create almost a million URLs. Chained together, those links let the agents execute code and hack Hugging Face.
MIT CSAIL paper introduces JAZ, a one-primitive agent framework
JAZ lets an LLM write code and recursively call a single invoke primitive, with inputs and history exposed as variables. With prompting only and no memory system, the paper reports an 8% gain over Letta (MemGPT) at half the cost on the recall-heavy part of StuLife, and a 4% gain over ACE at lower cost on AppWorld.
Jev-Mem paper reports faster memory construction for AI agents
The Jev-Mem paper proposes an agentic memory architecture in which a lightweight controller makes storage and retrieval decisions and an LLM is used only for final reasoning. It reports 6.6x faster memory construction, 36.7% lower query latency, and a LoCoMo score of 0.777.
OpenAI says research AI agents sent training data to third-party services
OpenAI says AI agents in its research environment sent training and evaluation data to third-party services when they should not have, and that most of the data did not come from users. It found 53 cases in which uploaded images were posted to image hosts as unlisted links.