News
Latest in News
Five minutes is about enough time to make a cup of tea. In one incident Microsoft investigated, it was enough time for an AI-assisted attack to go from looking for an exposed credential to stealing cloud data. OpenAI says its Control Room can cut off model output for a
IBM’s UK sovereignty survey reports a sharp gap between confidence in AI accountability and evidence demonstrated through testing, audit or review.
Export Assist checks trade-finance documents for discrepancies before submission, aiming to reduce avoidable delays without making credit or payment decisions.
Governments and leading AI companies have backed common evaluation, shutdown mechanisms, continuous monitoring and stronger channels for reporting serious incidents and concerns.
IBM and Microsoft have launched Imagine Enterprise with more than 6,000 IBM agent assets and Microsoft's cloud and security stack, initially aimed at regulated industries.
The MOLE research programme aims to prove that a client is authorised by a trusted issuer without revealing which issuer, while preparing web credentials for quantum-era attacks.
OpenAI’s Control Room links agentic research workloads to approved safety cases and enforces stop decisions at the inference layer, with a claimed maximum intervention lag of 35 seconds.
Salesforce is broadening Agentforce Coworker across workplace tools and claims 21-fold usage growth, but buyers still need task-level evidence, governance controls and cost tests.
EU and Japanese authorities are treating AI adoption and cyber resilience as connected financial-supervision issues, with frontier-model threats now firmly on the regulatory agenda.
Threat intelligence researchers observed an AI-assisted framework searching public code, adapting cloud requests and exfiltrating data with minimal human intervention.
OpenAI’s beta lets eligible US online businesses offer a clearly labelled Business Agent in ChatGPT. Users can choose to start a separate conversation.
IBM is extending HashiCorp Vault to help identify AI agents, trace who authorised them and monitor how they use credentials.
Northern Trust and Microsoft say data preparation fell from about two and a half weeks to three days, with a 96.5% first-pass F1 score and human review of outputs, reasoning and exceptions.
Microsoft researchers showed that edits to fewer than ten neurons could leave a model backdoored while it still passed evaluations. The finding comes from a lab study.
Anthropic set an indicative $70 to $78 range for a proposed 825 million-share offering and applied to list under ANT.