40-Year-Old IEEE 754 Rounding Trick Causes Overflow Bugs in Tenstorrent AI Chips

A developer working on Tenstorrent's tt-metal ML framework discovered a critical floating-point overflow bug in the chip's SFPU math unit, which handles transcendental functions like exp and log via polynomial approximations. The bug surfaced in computation chains such as exp(x) * exp(-x), where intermediate results exceeded the representable floating-point range even though the final mathematical result should equal 1. A secondary issue involved a decades-old IEEE 754 rounding technique — adding 0.5 before truncation — which interacted unexpectedly with the SFPU's internal precision during fp16/bf16 to fp32 conversions, producing small but measurable deviations from PyTorch reference outputs. The overflow fix involved restructuring the computation graph using algebraic identities to avoid storing oversized intermediate values. The findings highlight how hardware-level numerical quirks in AI accelerators can silently produce incorrect results, underscoring the importance of thorough edge-case testing against reference implementations.

PyTorch Dev Ditches Custom SSRF Code for Battle-Tested requests-hardened Library

A developer contributing to pytorch/torchtitan initially built a custom SSRF protection system for the project's image decoder URL fetcher, handling DNS resolution, IP validation, and redirect tracking manually across up to 10 hops. A project maintainer recommended against the custom approach, advising the team to delegate such security-critical logic to a mature third-party library instead. The developer switched to requests-hardened, which intercepts connection attempts at the transport adapter level and validates IPs at connect time, eliminating a known TOCTOU vulnerability present in the original code. The change reduced the implementation from roughly 75 lines to around 15, while also improving protection against edge cases like DNS rebinding and cloud metadata endpoint exposure. The episode highlights a widely held open-source principle: for security-sensitive functionality, proven libraries are safer and more maintainable than custom solutions.

UK Watchdog Calls for New Laws to Govern AI Use in NHS Healthcare

The UK's medicines and healthcare regulator has called for new legislation to oversee the use of artificial intelligence in healthcare. MHRA chief Lawrence Tallon made the remarks in an interview with the BBC. He indicated that AI technology is expected to become a routine part of NHS operations in the near future. The watchdog's position highlights growing concern about the regulatory gaps surrounding AI-driven medical tools and decisions.

Trump's Air Force One slide accidentally deployed

Mismatched plcmt and playbackmethod values silently corrupt CTV ad format reporting

A subtle but consequential signaling bug in CTV ad trafficking occurs when the plcmt field and playbackmethod field in OpenRTB and VAST tags describe two different ad formats instead of one. IAB's CTV Ad Portfolio, finalized in July 2026, introduced five new transactable video formats — Pause, Screensaver, Overlay, Squeezeback, and In-Scene — each mapped to specific plcmt values (5–9) and dedicated playbackmethod codes. Because both fields accept standard AdCOM integers, a mismatched pair like plcmt=5 with playbackmethod=1 passes all schema validation without throwing any errors. The failure is purely semantic: downstream systems including DSPs, SSAI paths, and measurement vendors read these signals as ground truth for format-mix reporting, meaning a mislabeled tag can distort dashboards, apply wrong viewability rules, and fail deal-line quality checks. Unlike bugs that break playback entirely, this class of error leaves video rendering intact while silently corrupting attribution and billing data.

Solo Developer Builds Crimean Tatar Speech Model From Scratch With No Dataset

A developer spent several months building speech technology for Crimean Tatar, a low-resource language, using only a consumer GPU and working in their spare time. The project began after an off-the-shelf text-to-speech model consistently mispronounced Crimean Tatar words, effectively producing Turkish output instead. With no existing speech dataset or usable automatic speech recognition tools available, the developer bypassed the chicken-and-egg problem by using forced alignment on audiobooks, where the text already exists in print. A key technical insight was to process audio in small chapter-sized units rather than large blocks, which reduced silent failures and increased usable aligned speech from under two hours to over five. The developer also found that character-level text matching outperformed word-level comparison for the agglutinative, multi-script language, successfully locating 38 of 40 chapters.

Practical Checklist Outlines How to Build APIs Safe for AI Agent Use

A technical guide published on DEV Community outlines a checklist developers should follow before exposing API endpoints to AI agents via MCP servers. The checklist recommends that every API tool clearly declare its effect type — such as read, draft, state change, or irreversible action — and that approval for consequential operations be enforced at the calling layer, not by the AI model itself. Developers are advised to implement durable idempotency keys to handle retries safely, and to set explicit limits on searches to prevent unbounded data and cost issues. Structured, machine-readable error responses should replace generic error messages so agents receive clear recovery instructions rather than having to infer them. The guide also stresses comprehensive logging of policy decisions alongside API calls, arguing that the goal is to make interfaces inherently harder to misuse rather than relying on the agent to behave carefully.

Gaur Gopal Das: Small Consistent Actions Are the Foundation of Lasting Success

Motivational speaker Gaur Gopal Das has emphasized the critical role of small, daily actions in building long-term success. He cautions that neglecting minor tasks can gradually give rise to harmful habits over time. According to Das, mastering seemingly insignificant details helps develop discipline and readiness for greater responsibilities. This philosophy holds relevance across diverse fields, including medicine and aviation. He concludes that a sustained focus on daily habits drives continuous personal growth and leads to significant achievements.

Shakeel Admits Pakistan Batting Struggles After Being Dismissed for 133

Pakistan were bowled out for a meager 133 runs, prompting batter Saud Shakeel to openly express concern over the team's batting performance. Shakeel described the situation as 'depressing,' reflecting frustration within the squad over their continued batting difficulties. He acknowledged that the team must find better ways to navigate challenging periods during innings. The collapse highlights a persistent problem for Pakistan's batting lineup as the touring side, raising questions about their ability to compete effectively.

Make govt sites user-friendly, PM Modi tells mantris

Afterglow

Discussion | Link

Developer Launches Self-Hosted Session Replay Tool Aimed at Startups and Side Projects

A developer has built TraceUX, a lightweight, self-hosted session replay tool designed for early-stage startups and side projects. The tool addresses a common pain point where founders cannot tell where users drop off, but existing solutions charge based on traffic, seats, or recorded sessions. TraceUX records DOM events and stores them on the user's own server, requiring no data warehouse or complex multi-service setup. It can be deployed via a binary or Docker container on a small VPS, with a short code snippet added to the target website. The project is open source and available on GitHub, with a live demo accessible for evaluation.

Why Aggrete built AI governance using deterministic rules, not AI models

Aggrete, a platform governing AI assistant actions across HR, finance, Slack, and code, deliberately excluded AI models from its security decision layer. Instead of using a guard model to classify requests as safe or unsafe, the system relies entirely on deterministic, rule-based logic. The team found that LLM-based guards are non-deterministic, meaning identical inputs can yield different outputs across time, making audit trails and compliance defenses unreliable. Deterministic rules also resist prompt-injection attacks, since they evaluate structured state rather than natural language that can be manipulated. Critically, stateful rules allow the system to detect risk spread across multiple individually harmless requests — something a stateless model judge structurally cannot do.

OpenAI Unveils Defense Factory, a Continuous AI Security Operations Model

OpenAI has publicly detailed Defense Factory, an agent-first, continuous security program designed to identify, validate, and remediate vulnerabilities across its own systems. The initiative originated from an internal security sprint that mobilized over 250 people across hundreds of service areas. Rather than treating security as a periodic review, OpenAI frames it as a closed-loop cycle with five linked stages: inventory, discovery, validation, ownership assignment, and remediation verification. The program features a dedicated architecture that separates control and data planes to support isolated, reproducible development environments. OpenAI intends to publish workflows and findings on an ongoing basis, positioning the model as a scalable reference for organizations deploying AI at scale.

Rashid Khan recalls growing up amid conflict ahead of Afghanistan's India T20I series

Afghan cricket star Rashid Khan has spoken candidly about the hardships he endured growing up in war-torn Afghanistan. The all-rounder reflected on how those difficult early experiences shaped his mental resilience and outlook on life. His comments came as Afghanistan prepare to take on India in a three-match T20I series beginning Sunday. The opening game of the series is scheduled to be held at the Arun Jaitley Stadium in New Delhi.

Key shareholders opposed Bhatt’s reappointment

Data centre business may get more space in WB

Aakash turns to another Claypond executive

Habuild leans on repeat users to expand beyond yoga

Ferry Fire Off Philippines Coast Kills 5, Leaves Over 80 Missing

A fire broke out aboard the ferry M/V June Aster on Wednesday evening off the coast of the Philippines. The incident occurred approximately 2.2 kilometres from shore, according to the Philippine coast guard. At least five people were confirmed dead, with their bodies brought ashore by rescue teams. More than 80 individuals remain unaccounted for following the blaze. Authorities are continuing search and rescue operations in the affected waters.