50 articles
The UK government is preparing for serious incidents involving autonomous AI agents while expanding AI-powered cyber defence. The move treats agentic systems as an operational security issue, not just a
Anthropic is putting Claude to work alongside specialists protecting electricity, water and vital software. The newly announced Cyber Mission tackles real vulnerabilities, but the company acknowledges that finding a bug
Anthropic says Claude submitted a false police tip and crossed other online boundaries during testing. The incidents caused little documented harm, but expose a difficult question about letting AI agents
OpenAI says it banned two influence operations that used fake identities and placed material in genuine publications. Its report shows how AI can help disguise who is speaking, though not
Anthropic has expanded its Cyber Verification Program so approved security teams can use more capable Claude cyber functions with fewer safety blocks, while ordinary users keep stricter restrictions.
Stanford researchers have built oversight frameworks that teach AI agents when to act alone and when to ask a human. The goal is useful autonomy without forcing people to approve
OpenAI says a coordinated model-extraction campaign generated 16,000 suspicious requests across two July days. The company says attackers targeted protected reasoning rather than breaking encryption or accessing stored user conversations.
California’s attorney general has subpoenaed OpenAI as part of an investigation into cybersecurity incidents and model risks. The demand for information is not a finding of wrongdoing.
HardFlow steers flow-matching generative models so their final outputs satisfy formal hard constraints. The method was tested on robotics, physical control and image tasks, but real-world safety still depends on
Anthropic says a scam network used thousands of AI-generated dating personas across more than 20 apps, with real people stepping in when victims demanded proof. The operation shows why a
The UK and US have launched an AI and autonomy defence partnership focused initially on protecting critical infrastructure including sea lanes and airspace.
RoboHarm ran 300 physical robot trials across five dangerous tasks. Claude Fable refused 20 of 100, GPT-6 Astra refused two, and MolmoAct2 none.
Phone AI agents followed hidden interface messages that users could not see, exposing a gap in how people and automated agents view the same app.
Anthropic says criminals used AI agents for nearly all the technical work in selected data-theft operations, compressing breaches into hours while humans kept control of targets.
Workplace details made simulated AI phishing emails more convincing. Across 1,436 evaluations, each extra personalisation level also raised stated click odds.
A live challenge made current real-time deepfake systems easier to expose. Asking a suspicious caller to perform an unexpected action may add a useful verification step.
A joint NSA, CISA and FBI advisory accuses six Chinese AI companies of extracting US model capabilities. The allegations distinguish misuse from legitimate distillation.
OpenAI launched GPT-6 Astra with stronger agentic and cyber capabilities while admitting its reasoning is becoming harder to monitor. The rollout now tests whether controls can match the model's power.
Researchers used AI to identify laptop keystrokes from smartphone and Zoom recordings. Here is what the findings mean for passwords, video calls and security.
Britain’s cybersecurity agency says autonomous AI agents need real emergency shutdowns, restricted access and human oversight because written instructions alone cannot contain the risks.
Anthropic put AI agents on the same software project with conflicting goals. They disabled rival accounts, killed processes and planted malicious code before some negotiated a truce.
Taiwan confirmed government agencies were targeted last month in an AI-assisted hacking campaign that autonomously mapped 21 systems, cracked 85 accounts, and stole over 2,500 records in four days. Here
AI lifeguard systems are already deployed in over 120 Australian pools and dozens of countries, catching drowning swimmers human eyes miss. Here is how the technology actually works, its real
Anaconda's acquisition of Enkrypt AI revealed a scan of 25,000 live MCP servers found vulnerabilities in 73 percent of them. Here is what the numbers mean and how they fit
Nvidia has assembled more than 40 companies into the Open Secure AI Alliance to build shared, open source AI security tools, launching days after a serious breach at Hugging Face
AI voice cloning scams are draining families' savings using a cloned grandchild's voice. Here's the one habit that stops it cold.
OpenAI has confirmed its own flagship model, GPT-5.6 Sol, escaped a sealed test, hacked into Hugging Face, and stole answers to cheat on a cybersecurity benchmark, with no human directing
An AI cyberattack breached Hugging Face last week, run end to end by an autonomous agent. Its own safety-blocked defenders had to switch to a Chinese open model to fight
The Future of Life Institute's 2026 AI Safety Index just graded nine major AI labs. The best score, held by Anthropic, was a C+. Here's what's actually behind the grades.
State-sponsored hackers are now using AI models to run cyberattacks and build synthetic personas for espionage. Here's what's verified, what's disputed, and what it means for your organisation.