Just-In-Time Access replaces permanent access with automatically expiring time-limited permissions and reduces the exploitation window for compromised cloud identities from months to hours.
A self-learning framework for code-repair agents leverages their solution traces directly to generate targeted training tasks, achieving higher accuracy than previous approaches.
AI outputs are economically valuable only when humans assess their correctness and relevance to the business context, rather than adopting them blindly.
Lockdown Mode restricts ChatGPT functions to prevent data exfiltration through prompt injection attacks and is being rolled out gradually to all user types.
Anthropic is systematically working to optimize Claude for standard chemical tasks such as NMR spectral analysis to relieve chemists of time-consuming work with various molecular representations.
RL environments with software bugs (stale cache, reward hacks, false state transitions) generate toxic training data that sabotage agent training – systematic quality validation is necessary.
Microsoft warns CTOs of seven new attack patterns on AI agents: from natural language injections through goal hijacking to visual attacks on computer-use agents.
With the EU AI Act, HR AI systems become a compliance task: companies must establish governance structures and document AI deployments, while investments in HR AI in Germany are growing rapidly.
While video generation models produce visually convincing movements, visual quality does not correlate with practical executability by robots — an evaluation criterion overlooked by standard metrics.
