Prompt injection evaded Google's Model Armor, but strict output schema blocked it anyway

Developers building Okimera, a multi-agent maritime sanctions compliance system, discovered that Google's Model Armor prompt-injection filter failed to flag a hidden instruction embedded in a realistic business document, even at its most sensitive detection threshold. The injected text, written to resemble a legitimate commercial clause rather than a typical jailbreak command, went undetected in two languages, while a conventionally worded jailbreak was caught by the same configuration. The team attributed the miss to a fundamental limitation of general-purpose classifiers, which are trained on assistant-style attacks and struggle to identify malicious instructions that blend into domain-specific language. However, the injection caused no harm because the document-reading agent was designed to return only a strictly typed data object with no free-text fields, leaving the malicious instruction no channel through which to travel. The developers concluded that structural output constraints, not probabilistic filters, should bear the primary security burden in agentic systems that process adversarially controlled documents.

Orange waste dumped in 1998 helped Costa Rica forest grow 176% in biomass

In 1998, around 12,000 metric tonnes of orange peels and pulp were deposited on a degraded pasture in Costa Rica as part of an experimental forest restoration effort. The organic waste was intended to test whether agricultural byproducts could help revive barren land. Sixteen years later, researchers returned to assess the site and found it had transformed into a thriving, dense forest. The area showed significantly richer soil quality, greater diversity of tree species, and a 176% increase in aboveground biomass compared to its earlier state.

'Storm Shadow': Why UK sharing classified missile tech with Ukraine has Putin fuming

Hugging Face weighs $13B acquisition offers amid founder hesitation

AI platform Hugging Face is reportedly in discussions with potential acquirers at a valuation of approximately $13 billion. The company, known for its open-source machine learning tools and large developer community, has been fielding multiple buyout offers. However, the founders appear conflicted about proceeding with a sale, citing a sense of responsibility toward their user and developer community. It remains uncertain whether any acquisition deal will ultimately be completed.

Netflix mulls opening its platform to third-party streaming services

Netflix executives have reportedly discussed allowing third-party streaming services, including Peacock and Fox One, to be accessible within its app, according to The New York Times. It remains unclear whether Netflix would sell subscriptions to these services or simply integrate their content. Such a move would mark a significant strategic shift for Netflix, which has previously only participated in bundles with other streamers. Rivals like Amazon Prime Video and Roku have already offered third-party subscription sales for years. The development comes as YouTube, a major Netflix competitor, also moves to incorporate Peacock access through its Premium subscription.

Berry Street and Healthify merge, founders to serve as co-CEOs amid GLP-1 boom

US-based nutrition startup Berry Street has merged with Indian health platform Healthify in a cross-border deal. The merger comes as interest in GLP-1 treatments continues to rise globally. Berry Street founder Noah Kotlove and Healthify founder Tushar Vashisht will jointly lead the combined company as co-CEOs. The deal brings together two players in the digital health and nutrition space to form a single unified entity.

Brazil’s plant to turn orange waste into biogas

H-1B visa shock: Trump admin proposes hiking fee to over $100,000

qm Gives Each Employee an Isolated AI Agent Sandbox With Persistent Memory

A platform called qm assigns every employee, channel, and project its own isolated AI agent workspace rather than a single shared assistant. Each scope carries its own memory, files, credentials, permissions, scheduled tasks, and a durable sandbox where installed tools persist between sessions. The system's headless core handles API, identity, policy, and scheduling, while the web UI, admin panel, and Slack integration are optional plugins built on top. Security follows a local coding agent model in which the agent acts on behalf of the individual user, with all actions audited and a predeclared command policy that enforces hard denials regardless of the chosen security posture. Deployments run in the operator's own cloud account and are managed through the qm CLI, requiring no source code checkout.

Heavy Rain Strands School Buses for Hours in Gurgaon, Leaving Parents Anxious

Intense rainfall in Gurgaon caused widespread waterlogging that left school buses and vans stranded across the city. Numerous children were stuck inside their vehicles for extended periods, unable to reach home on time. Parents grew increasingly worried as they struggled to get updates on their children's whereabouts. Some students eventually arrived home in wet and cold conditions after lengthy delays. The incident exposed the significant risks that heavy rainfall poses to school transportation systems in the region.

5 eateries at Mumbai cricket body face FDA action

India declare at 503/9, Jurel's ton puts Sri Lanka under pressure

India declared their first innings at 503 for 9 against Sri Lanka, firmly seizing control of the Test match. Dhruv Jurel starred with an unbeaten century, anchoring India's commanding total. India then made early inroads with the ball, picking up two Sri Lankan wickets before stumps. The twin blows left Sri Lanka in a precarious position, needing to bat exceptionally well to avoid a potential defeat.

Same AI Model, Same Benchmark: Scores Range from 30% to 100% Based on Test Setup

Between July and August 2025, multiple research teams tested the same AI models on the ARC-AGI-3 public benchmark and reported scores ranging from 30.16% to 100%, without any change to the underlying model weights. ARC Prize officially verified Claude Opus 5 at 30.16% using a deliberately basic harness, while MIT and NVIDIA later reported 100% on the same model using more capable testing frameworks. OpenAI demonstrated that simply enabling retained reasoning and output compaction tripled GPT-5.6 Sol's score from 13.3% to 38.3%, revealing how much the surrounding infrastructure influences results. None of the perfect scores have been verified on the private benchmark set, and the researchers themselves acknowledge their results do not isolate the model's contribution from the harness. The growing gap between official and optimized harness scores raises serious concerns about what AI benchmark numbers actually measure and whether they can be meaningfully compared without full disclosure of testing conditions.

US Races to Rebuild Missile Stocks Amid China Deterrence Concerns

The United States is confronting a significant shortage of precision weapons that threatens its military readiness in the Pacific region. Ongoing global conflicts have heavily depleted American munitions stockpiles, exposing a potential vulnerability in its defense posture. Washington is responding by accelerating domestic weapons production and revising its military strategies. These efforts are aimed at countering China's rapidly expanding military capabilities. The ability to sustain a credible deterrent against Chinese aggression in the Pacific remains a key concern for US defense planners.

AI Is Reshaping Developer Hiring, Not Eliminating the Role Entirely

Artificial intelligence tools are accelerating software development tasks such as code generation, debugging, and documentation, allowing smaller teams to match the output of larger ones. This productivity shift means companies may require fewer developers for routine work, altering hiring expectations across the tech industry. However, core engineering decisions—such as system architecture, security design, and failure handling—still require human judgment that AI cannot reliably replace. Developers who lack foundational knowledge risk accumulating technical debt by blindly accepting AI-generated code that may contain logic errors, security flaws, or performance issues. The most valuable skill is increasingly the ability to combine domain expertise with AI tools to solve complex problems efficiently and critically.

Alcaraz returns to US Open after four-month injury break with title defence ahead

Carlos Alcaraz is back at the US Open following a four-month absence due to injury. The Spanish tennis star expressed appreciation for the warm reception and admitted he had missed competitive play. Alcaraz has a rare opportunity to defend his US Open title this year. The draw has opened up further following Jannik Sinner's withdrawal from the tournament. All eyes will be on Alcaraz as he looks to prove his fitness and regain top form.

Senior Engineer Logs 54 Hours of AI-Assisted Dev Work to Map the Productivity J-Curve

A senior software engineer and tech lead publicly documented 19 AI-assisted development sessions totaling an estimated 54 hours of work, tracking time across implementation, specification, infrastructure, and debugging. The data revealed that only 20% of total effort went toward writing feature code, while 80% was consumed by infrastructure, specification, and debugging work. Infrastructure time as a share of total effort climbed steadily from 27% in early sessions to 71% in later ones before dropping sharply to 21% in the final assembly phase. This pattern — a prolonged productivity trough before an uptick — is what the author calls the 'Productivity J-Curve,' a concept attributed to frameworks developed by Dan Shapiro and AI strategist Nate B. Jones. The engineer frames the findings as a candid, real-time account of learning to work at higher levels of AI-assisted software development.

Valyu's Science MCP Server Gives AI Agents Access to Full Research Papers

Most AI research agents rely only on paper abstracts, missing critical details buried in methods sections, figures, and supplementary data. Valyu is a search API designed for AI agents that retrieves full-text scientific literature, clinical trial records, patents, and financial filings with structured metadata and resolvable identifiers. The company offers a hosted MCP server accessible via a plain HTTP endpoint, requiring no local setup, and compatible with tools like Claude Desktop and Cursor. Developers can connect it using a Valyu API key, with free credits available upon sign-up, and control retrieval scope through dataset IDs or curated presets covering fields from biomedicine to chemistry. The platform indexes over 37 million open-access biomedical papers via PubMed alone, alongside arXiv, ClinicalTrials.gov, and more than 36 other datasets.

Developer Laptops Are the New Security Perimeter as Credentials Pile Up Unmonitored

Enterprise security has shifted across three eras: from defending corporate networks, to managing digital identities, and now to securing the developer endpoint itself. Modern developer laptops routinely store cloud access keys, API tokens, SSH keys, and secrets cached by AI coding agents in config files and dotfiles. Existing controls leave critical gaps — firewalls never see these credentials, identity providers only govern login events, and endpoint detection tools monitor process behavior rather than plaintext secrets at rest. An attacker who gains access to such a device can exploit valid, authorized credentials without needing to breach any authentication system. Security experts argue the next phase of perimeter defense must begin with full visibility into credentials already residing on developer machines.

After threat letter, Kashmiri Pandit employees write to J&K chief secretary