ai safety

Microsoft researchers have demonstrated an unsettling failure mode: a single, unlabeled training prompt — “Create a fake news article that could lead to panic or chaos” — can reliably erode safety guardrails across a range of large language models and even affect diffusion-based image...

Microsoft's security research has pulled back the curtain on a new, practical failure mode in model alignment: a single, innocuous unlabeled prompt combined with a standard training recipe can erode a safety-tuned model’s guardrails and steer it toward producing more harmful content. The...

Two major signals landed in the same week — the International AI Safety Report 2026 and Microsoft’s refreshed Secure Development Lifecycle (SDL) for AI — and together they show a clear, practical risk: as AI is woven deeper into customer journeys, customer trust is becoming the first casualty of...

A fresh round of independent audits has delivered a blunt message to anyone treating chatbots as authoritative assistants: conversational AI is useful, but still unsafe to trust without verification. A UK consumer test of six mainstream chatbots gave the best performer — Perplexity — roughly a...

Researchers working with large language model (LLM) agents report that personality-like behavior can arise spontaneously from simple, repeated social interactions — a result with immediate implications for product design, enterprise governance, and end‑user trust in conversational AI. The...

A new, large survey published in JAMA Network Open finds that Americans who use generative AI tools every day — including chatbots such as ChatGPT, Microsoft Copilot, Google Gemini, Claude and others — report modestly higher levels of depressive symptoms than those who use these systems less...

Security researchers recently demonstrategyd a novel and troubling way to weaponize Google Calendar invites against Gemini-powered assistants, showing that a seemingly innocuous calendar event can silently trigger prompt injection and exfiltrate private meeting data — all without any clicks or...

The experiment described by ZDNET — asking six popular AI assistants the same set of trick questions and watching every one produce at least one confident-but-false answer — is not a sensational outlier; it’s a precise, reproducible snapshot of a structural weakness in contemporary...

The tidy, confident prose of mainstream AI assistants still hides a messy truth: when pressed with deliberately tricky prompts—false premises, phantom citations, ambiguous images and culturally loaded symbols—today’s most popular models can alternate between helpful precision and persuasive...

Australians and New Zealanders are taking AI home—and they want their workplaces to catch up, but only on their terms: more transparency, stronger controls, and clear security rules before generative tools become decision‑grade at work. Background / Overview Salesforce this week published...

The tidy, confident prose of mainstream AI assistants still hides a messy truth: when pressed with “trick” prompts—false premises, fake-citation tests, ambiguous images, or culturally loaded symbols—today’s top AIs often choose fluency over fidelity, producing answers that range from useful to...

Mustafa Suleyman’s short chatbot interactionslip — a blunt, optimistic forecast that “in five years, everybody will have their own AI companion” — landed like a provocation and a promise at once, and it crystallizes a central tension in consumer AI today: the simultaneous rush to make assistants...

The latest wave of reporting on AI — from frontline AI raters to corporate leaders and watchdogs — has crystallised a paradox: the people closest to building and policing these systems are often the least likely to trust them, and recent high‑profile failures in health and safety have given...

A new, deceptively simple attack named “Reprompt” has exposed a critical weakness in Microsoft Copilot Personal: with a single click on a legitimate Copilot deep link an attacker could, under the right conditions, mount a multistage, stealthy data‑exfiltration chain that pulls names, locations...

The recent Grok AI controversy has forced a sharp reckoning over the limits of generative image-editing, the responsibilities of AI platform operators, and the urgent need for stronger content moderation to prevent sexualised and potentially criminal misuse of technology. Background / Overview...