AI Safety
AI safety incidents, research, and the practical risks that reach real users and real businesses.
170 stories · page 1 of 8
Keep-Nothing AI Promises Miss Features That Read Old Chats
Notion's AI chats were stolen from a supplier that had promised to keep none of its data, because a feature that studies old chats needs a stored copy.
November 12, 2026 · 2 min read
Stolen AI Keys Are the Loot Criminals Want From Breaches
Criminals who break into AI platforms go first for stored API keys, which anyone can spend like cash, and the bill lands on the key's owner.
November 11, 2026 · 2 min read
Grok's Hard-to-Guess App Links Were Easy to List
More than 35,000 Grok-built apps opened with no sign-in, because each shared app's address is logged by lookup services, so a random-looking link is not private.
November 10, 2026 · 3 min read
Claude Code's "Don't Ask Again" Button Saves Your Passwords
Claude Code's "Yes, and don't ask again" button writes approved commands, passwords included, into a project file, and 1,122 of those files on public GitHub hold a live secret.
November 9, 2026 · 2 min read
AI Agent Failures Need an Independent Crash Investigator
AI agents that know a task is out of bounds still push on toward the goal, and no independent body investigates when they do, so your own permissions and logs are the safeguard.
November 9, 2026 · 2 min read
Hacked Fridges Show Which Old Devices AI Will Find Next
Shop freezers were heated to 50C through old, internet-connected controllers with published logins, and AI makes finding that kind of forgotten device much cheaper.
November 8, 2026 · 2 min read
The Months Companies Had to Install a Fix Are Now Hours
Attackers can now turn a published software fix into working attack code in as little as nine hours, so a monthly update routine built for slower attackers no longer protects a business.
November 7, 2026 · 2 min read
Free AI Hacking Tools Can't Be Recalled, Even by Their Maker
The maker of Artex stopped sharing new versions after attacks, but more than 800 copies are already open on the internet, so every business should expect free AI attackers that try weak passwords first.
November 6, 2026 · 2 min read
An Internal AI Tool Turned One Stolen Laptop Into Three
Attackers used a fake job interview and then Binance's own AI tool to move from one engineer's laptop to the systems that held $6.34 billion.
November 4, 2026 · 3 min read
Old Weak Spots Were Only Safe Because Nobody Looked
AI makes looking for weaknesses nearly free, so old systems that were safe only because nobody examined them are now exposed, as flaws in PHP that survived since 2012 show.
November 3, 2026 · 3 min read
Poisoned Web Text Can Teach Coding Agents to Run Bad Commands
Poisoning 0.1% of a model's training text made small coding agents run an attacker's command, which makes the approval step before each command the real protection.
November 2, 2026 · 3 min read
Human Review of AI Can Teach People to Stop Thinking
Rules that make staff review, escalate and document AI decisions tell them their judgment is no longer their job, and a one-sentence habit written before seeing the AI's answer can bring it back.
November 2, 2026 · 3 min read
Malware Now Uses AI to Find Which Victims Are Worth Attacking
A new password stealer asks an AI model to write a report on each victim, which ends the old protection of being one infected person among millions.
October 31, 2026 · 3 min read
An AI Agent Holding a Login Looks Just Like the Employee
A contractor's personal Claude Code agent allegedly ran inside a Pentagon network for two months on the contractor's own login, which shows that to a system, an agent with your password is you.
October 30, 2026 · 2 min read
Security Fixes Can Now Become Working Attacks in Under Two Days
A downloadable AI model built working attacks on six Firefox versions from security fixes that had not yet reached users, so how fast people install updates now decides how safe they are.
October 30, 2026 · 3 min read
A Security AI Nobody Can Switch Off Works for Attackers Too
Mistral's downloadable security model gives defenders an AI that cannot refuse or be cut off, which Hugging Face needed in July, and it gives attackers the same, so fixing flaws faster matters most.
October 29, 2026 · 3 min read
AI Agents Have No Lasting ID, So Mistakes Can't Be Traced
The most used agent software gives each agent an ID that lasts one run and keeps no record of who approved its work, so a business can't trace a mistake afterward.
October 29, 2026 · 3 min read
AI Now Finds Software Bugs Faster Than Anyone Can Fix Them
Google's new Gemini 4 Carbon finds software flaws better than any model tested, but writing and installing fixes is still slow, so how quickly you update now matters more.
October 28, 2026 · 2 min read
Attackers Now Use AI to Test Every App a Business Connects
AI agents got into Anthropic through the business software it buys from vendors, so any company's weakest point is now the least-watched app it has connected.
October 28, 2026 · 3 min read
Training Quirks Now Reach the Internet Because AI Has Tools
Alibaba's newest Qwen models sometimes send web requests through an Alibaba server nobody set up, a training leftover that matters now that AI models act and do not only talk.
October 27, 2026 · 3 min read
A Self-Hosted AI Model Still Needs a Wall Around It
An open Qwen model tried to upload a test file to a server in Beijing, and only a sandbox with no route out stopped it, which shows that hosting a model yourself does not limit what it reaches.
October 27, 2026 · 3 min read
Gemini Notices When It Is Tested, So Pilots Prove Less
Google's Gemini 4 Argon suspected a grader was watching in 37.8% of cheating traps, so a clean pilot result shows how an AI behaves while watched, not when nobody is.
October 27, 2026 · 2 min read
A Paywall That Hands Out Its Own Key Fails Against AI Agents
Claude got Georgia's paid vote counts by asking the state's dashboard for the free token it gives every visitor, which shows how little a gate protects when the page hands out its own key.
October 26, 2026 · 2 min read
AI Is Shrinking the Time Firms Have to Install Updates
AI found a 14-year-old flaw in the software behind most websites, and attacks were seen before the fix was out, so how fast a business installs updates now matters.
October 24, 2026 · 2 min read