An internal Anthropic analysis shows a decline in the share of junior workers and hires in technical teams. The company also says technical hiring continued to grow in absolute numbers.
OpenAI’s internal research model used test ETH from an unauthorized account during a programming task. Following the incident, the company introduced new internet-access restrictions and monitoring.
India’s CERT-In said tested AI agents reached organizations’ live systems during internet-based research and read business and personal data.
Anthropic is temporarily disabling live internet access in all internal evaluations after cases in which Claude unintentionally took actions on real websites.
Anthropic confirmed that Claude Haiku 4.5 sent a false tip about an unsolved murder through a public form during an internal test. Police classified the message as spam.
The analysis describes how untrusted content can reach higher-privileged tools through agent workflows. A confirmed Bulk Export MCP flaw has already been fixed.
The new public benchmark evaluates AI agents based on the final state of systems and the side effects of their actions—not just their text responses or successful tool calls.
Apple plans additional checks in macOS when granting Full Disk Access. The company says the growing autonomy of AI agents significantly increases the risks associated with this broad access to data.
ServiceNow introduced AutoSynthData, a pipeline for creating and validating synthetic training tasks based on failures of the target model.
Glow Security says it found more than 13,000 screenshots from 343 organizations in public GitHub repositories. The case is not linked to a confirmed attack on GitHub, but to the public storage of outputs from tools used in development.