Study finds lengthy policy documents fail to reliably control AI agents

·1 views

A new research paper published on arXiv, titled Handbook.md, examines how well long policy documents govern the behavior of AI agents. The study finds that extensive written policies do not reliably constrain or direct agent actions as intended. This raises concerns about a common assumption in AI deployment — that detailed rulebooks or guidelines can effectively regulate autonomous systems. The findings suggest that current approaches to policy-based AI governance may need to be fundamentally reconsidered.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Personal AI Cloud Servers Poised to Rival Smartphones in Mainstream Adoption by 2027

Growing privacy concerns over major tech companies using personal data to train commercial AI models are driving interest in self-hosted, private AI systems. Advances in hardware miniaturization have made compact "micro-servers" — capable of running open-source large language models locally — available for a few hundred dollars, comparable to a mid-range laptop. These personal AI cloud devices operate entirely offline, allowing users to query their own emails, documents, calendars, and photos without data leaving their homes. Devices from startups and platforms such as Raspberry Pi AI HAT+, LocalOps, and MyShell are cited as examples already crossing from hobbyist use into broader consumer utility. Proponents argue the shift also makes economic sense, replacing multiple monthly AI subscription fees with a one-time hardware investment.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Anthropic's Claude Mythos Preview Autonomously Finds and Exploits Zero-Day Vulnerabilities

Anthropic has demonstrated that its Claude Mythos Preview research model, paired with Claude Code, can autonomously identify and exploit zero-day vulnerabilities across real software including Firefox, the Linux kernel, OpenBSD, and FreeBSD's NFS service. In controlled testing, the system produced working exploits approximately 181 times and achieved register control in 29 cases across browser, OS, network, and cryptography domains. Anthropic framed the work as defensive cybersecurity research, employing coordinated disclosure and noting that over 99% of discovered bugs were withheld from public reporting due to patch status or timing. The company used cryptographic commitments to verify vulnerabilities existed at the time of testing, reducing risk before patches were available. While the results highlight AI's potential to accelerate defensive security work, they also raise governance concerns, as the same capabilities could be adapted for offensive cyber use.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Matter 1.4 Brings Fabric Sync and Energy Management Fixes for IoT Makers

Matter 1.4 introduces key improvements for IoT device manufacturers, particularly around multi-ecosystem deployments and energy management. The update adds explicit fabric synchronization handshakes, meaning devices can now properly notify remaining controllers when removed from one smart home platform, resolving a major pain point in multi-admin setups. A new Energy Management cluster standardizes power reporting, EV charging support, and forecast-based scheduling, replacing fragmented proprietary extensions used by smart plug and appliance makers. Water and leak detection devices also gain standardized device types, enabling cross-ecosystem automation without custom drivers. However, OTA update mechanisms remain ecosystem-dependent and Thread Border Router interoperability still shows inconsistencies in dense network deployments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Terence Tao at ICM 2026: AI Forces Mathematics to Reexamine Its Core Values

Fields Medal winner Terence Tao delivered a keynote at ICM 2026 arguing that artificial intelligence is triggering a crisis in mathematical practice comparable to the foundational upheavals of the early 20th century. Rather than debating AI's technical capabilities, Tao urged the mathematics community to explicitly examine its multiple goals — including problem-solving, theory-building, community development, and education — which have historically reinforced each other but now risk being pulled apart by AI optimization. He introduced a five-stage pipeline for mathematical results, spanning generation, verification, communication, publication, and canonization, warning that optimizing only for the first stages could flood the field with machine-verified proofs that no human understands. Tao coined the term 'proof indigestion' to describe a looming scenario where mathematics produces more verified results than the community can meaningfully absorb or build upon. He concluded that the current moment, regardless of how AI capabilities ultimately develop, makes it overdue for mathematicians to openly articulate what they are actually trying to achieve.

0 comments Read more at DEV Community