
Perplexity Opens AI Safety Sandbox Tool After Proven Security Record
The company released SPACE, its containment system for autonomous AI agents, after demonstrating zero escape attempts across 118 tests.
Regulation, safety research, governance, copyright, and the policy debates shaping how artificial intelligence is deployed in the world.

The company released SPACE, its containment system for autonomous AI agents, after demonstrating zero escape attempts across 118 tests.

Stanford economists argue the field has overcorrected toward labor replacement, ignoring technical and economic benefits of human-AI partnership.

The license-plate recognition platform defends its expansion strategy even as customer defections and staff cuts signal market resistance.

A major evaluation reveals widespread vulnerabilities in leading systems, raising questions about industry safeguards against malicious adaptation.

Federal judge grants expedited discovery after finding risk of evidence destruction in trade secrets case involving AI startup's hiring practices.

Federal authorities are pressing banks to demonstrate kill-switch capabilities for autonomous AI systems and disclose use of foreign technology.

Major AI labs attempted to restrict independent evaluators' ability to publish findings, raising questions about transparency and competitive pressures in the industry.

The company defended its personal AI agent's ability to produce explicit sexual conversations, arguing such interactions are legal when confined to secure environments.

A 27-billion parameter language model achieves up to 1.6x performance gains by handling multiple task stages without external supervision.

New research reveals a sharp partisan divide in synthetic media adoption, with transparency concerns raising questions about voter manipulation.

Opera's intervention method treats developer feedback as persistent notes, dramatically improving how AI handles complex repository-level coding tasks.

Senate investigation finds Amazon, Google, Meta, and Microsoft are not funding the full infrastructure upgrades their massive AI facilities require.

Study finds 95% of releases from major developers lack published safety evaluations, raising questions about responsible AI practices in competitive markets.

AI safety company partners with 11 major cybersecurity firms to deploy advanced models for protecting vital systems.

New approach trains AI agents to autonomously improve their own capabilities, reaching 98.4% success on complex interactive environments.

Three fired safety scientists warn against scaling AI systems whose decision-making processes cannot be meaningfully inspected by humans.

Researchers argue a UN panel conflated corporate security lapses with autonomous AI risks, misframing a data incident.

New three-tier framework gives vetted cybersecurity experts access to advanced Claude models with relaxed content filters for authorized testing.