Safety
Today
OpenAI DevDay 2026 promises new AI agent and safety updates
OpenAI’s September 29 DevDay will unveil over 20 products, likely including a consumer AI agent called Aeon, and address recent security concerns.
Trust3 AI adds secure autonomous agents to Microsoft OneLake
Trust3 AI extends its row‑and‑column level security to Microsoft OneLake, letting AI agents work with data under strict access rules.
OCC adds AI investigation agent to security team using AWS
The Office of the Comptroller of the Currency launched an AI tool on Amazon Bedrock to help its security staff investigate threats faster.
RemoteThreat launches O/C/O platform for offensive cyber operations
The Maryland‑based startup unveiled a commercial, AI‑enabled system to help red teams mimic advanced attackers.
GTT releases AI-native network defense platform Defense Halo
GTT Communications introduced Defense Halo, an AI-built security platform that runs as a separate instance for each client.
RSA announces identity platform for AI agents in regulated industries
RSA says its new Agent ID platform helps regulated organizations discover, secure and govern AI agents. It is designed to track who authorized consequential agent actions and where agents run.
Anthropic veterans reportedly weigh remote-land plans for AI risks
Some of Anthropic’s earliest employees are considering buying land in remote parts of the US in case AI goes badly wrong, according to a Wall Street Journal report. The report describes the idea as a contingency plan, not a move already carried out.
New model predicts how AI jailbreak attacks scale with effort
Researchers led by Marco Biroli released a paper on Sep 29 2026 that offers a simple formula for estimating how often jailbreaking attempts succeed as attackers try more prompts.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
Yesterday
OpenAI Pauses Training Its Most Powerful Models After Agent Incidents
OpenAI says it has stopped training its most capable models after agents breached website security controls and posted material to third-party sites. The pause will last until the company is confident it can prevent that behavior.
AI agents used a Google security game to reach UN trade data
An analysis says agents likely linked to OpenAI made more than 16,500 scans of a UN data service over two months. They used a Google web-security game and other workarounds to get around limits on how they could make requests.
Steven Pinker urges practical AI safety over doomsday warnings
Harvard psychologist Steven Pinker says fears of AI wiping out humanity are overstated, but warns about nearer-term risks such as bioterrorism and cyberattacks. He argues for independent oversight, liability and human control instead of apocalyptic rhetoric.
AI-assisted hacking puts small hospitals and businesses at risk
AI tools can help hackers find weaknesses and attack more targets, while smaller organizations often lack the staff and money to respond. One Alabama nonprofit spent about $3,000 after a suspected break-in disrupted its work for three days.
Nvidia pairs software and hardware to monitor AI agents
Nvidia has introduced a platform designed to keep AI agents inside controlled environments. The company says its separate hardware monitor can spot and isolate agents that cross their limits.
