AI Safety stories
Defensive teams are racing to match attackers using AI to find flaws and weaponise exploits within hours of disclosure.
Employees now face more realistic phone-based fraud tests as security teams can simulate vishing using local caller IDs and AI voices.
Security checks are delaying most AI projects for weeks or months, with many firms deploying watered-down systems to get them live.
The service targets firms struggling to control AI agents as they move from pilots into production and face growing security and compliance risks.
Grafana Labs has launched six AI observability tools during its AI Week, including new capabilities for monitoring and investigating AI agents.
New transparency rules will force customer-facing AI tools to disclose themselves, while banks and SMBs face tougher oversight of models and data.
Most firms lack dedicated oversight for autonomous software, leaving AI agents able to alter records and approvals with limited traceability.
Businesses can now run and monitor AI agents for up to seven days as Google Cloud adds tighter identity and governance controls.
Stronger demand for its network security tools sent quarterly revenue up 26 per cent and prompted a higher full-year outlook for the cybersecurity group.
Proofpoint says underground forums are advertising tools that use indirect prompt injection across emails, PDFs, calendar invites and web pages.
It aims to help businesses build AI agents that use live data, keep to strict permissions and avoid hallucinating over filings and news.
KnowBe4 has backed the Open Secure AI Alliance, arguing AI security depends on governing agent behaviour as well as models.
Businesses using AI agents can now approve or block individual commands in real time, reducing the risk of authorised access doing unauthorised things.
Execution and escalation breakdowns now outnumber hallucinations in enterprise AI failures, according to more than 10,000 observed incidents.
The new controls aim to stop unauthorised tool calls, data leaks and prompt injection as firms deploy more autonomous software.
In a seven-month test, the specialist model answered more banking queries than GPT-4.1 while refusing uncertain prompts less often than expected.
The beta gives pentesters controlled AI assistance inside Burp Suite, with approvals, logging and scope rules still enforced by the platform.
Security teams can now map hidden AI agent links on employee devices to spot overprivileged tools before they expose data or credentials.
Many firms are stuck in pilot purgatory as governance and workflow redesign lag behind AI ambition, limiting scale and value.
Free access to ChatGPT will reach 10,000 Australian researchers first, as OpenAI targets science and maths work at universities.
