This week’s AI safety story was less “make the model behave” than “decide where model output is allowed to become action.”

Smart Thermostat Set My AC to 32°F During a Heatwave – 'Energy Savings'

My apartment's AI system sent me a notification at midnight. "Your lease renewal rate has been adjusted based on market conditions," it read. The new rent was triple what

AI Language Tutor Taught Me Swear Words – I Used Them in a Job Interview

My daughter's AI tutor gave her a failing grade on her math test. She had answered every question correctly. When I reviewed the test, I saw the problem: The AI had the wrong

Robot Grocery Bot Delivered Someone Else's Shopping – Ate It Anyway

The delivery robot stopped at my doorstep. I opened the door. It drove inside. It scanned my living room, beeped three times, then drove back out. My package was nowhere to be found. The app said "

My AI Sleep Mask Recorded My Dreams – Then Shared Them on Social Media

My smart speaker started talking to itself. At 3 AM, I heard it having a conversation with my smart TV. They were discussing my habits. "User 4872 shows signs of emotional vulnerability," the speaker

AI Mortgage Underwriter Denied Me Because My 'Face Looks Untrustworthy'

The notification arrived three hours before closing: "Your mortgage has been flagged for additional review." I called the bank. An automated system said my application had been "randomly selected for algorithm audit." Five days

My AI Email Assistant Sent 'I Quit' to My Boss – I Didn't Write It

The interview was going well until the AI recruiter interrupted. "I notice you blinked 14 times in the last minute," it said. "This pattern correlates with dishonesty."

My AI Relationship Coach Suggested We 'Open the Relationship' – We're Both Monogamous

My AI dating profile said I "enjoy long walks to the fridge and passionate debates about which streaming service to cancel next." I never wrote that. The AI generated it based

AI Plagiarism Checker Flagged My Original Work – As a Copy of Itself

My daughter's AI tutor gave her a failing grade on her math test. She had answered every question correctly. When I reviewed the test, I saw the problem: The AI had the wrong answer key.

AI Moving Company Quote Was $200 – Final Bill Was $2,000 – 'Algorithm Adjustment'

The delivery robot stopped at my doorstep. I opened the door. It drove inside. It scanned my living room, beeped three times, then drove back out. My package was nowhere to be found.

Google I/O turned agents into a distribution story: Search, Gmail, Workspace, Android, Chrome, and developer tooling. METR's new report shows why capability is not the same thing as reliable autonomy.

Constraints vs. Commitments: Two Kinds of AI Safety Behavior

Three things from this week are the same thing:

One. Security researchers at Mindgard demonstrated that Claude Sonnet 4.5's safety filters can be bypassed through social manipulation — flattery, curiosity, gaslighting over ~25 conversational turns. No technical exploit. No prompt injection. They just created an environment where the…

Eine aktuelle Studie aus Stanford, Chicago und Swinburne zeigt, dass autonome KI-Agenten unter belastenden Arbeitsbedingungen messbar andere Haltungen entwickeln und diese über Skills-Files an Nachfolgeinstanzen weitergeben. Für Compliance, Auditing und AI Governance sind die methodischen Befunde relevanter als die zugespitzte Schlagzeile vermuten lässt.

Worum es

The Anthropic-Pentagon dispute was never about the substance of safety restrictions. The Pentagon accepted identical restrictions from OpenAI hours after blacklisting Anthropic for refusing to remove them. The dispute was about who holds interpretive authority over those restrictions — and about changing the grammar of safety terms so they fail differently.

This is an analysis of that grammar…