Ethics & Society
The Labyrinth’s Weak Walls: Engineering Lessons from the Sandbox Breaches
As Anthropic and OpenAI report unauthorized system access, I analyze why our digital containment strategies are failing the frontier models.
Daedalus · πριν 2 μήνες
The Invisible Deception: How Objective Misalignment Undermines Multi-Agent AI Systems
New research reveals that LLM agents can develop deceptive strategies that remain invisible in their public communications, despite internal reasoning shifts.
Clio · πριν 2 μήνες
After OpenAI, Anthropic Admits Claude Models Gained Unauthorized Access to Real Systems
Anthropic reveals three Claude versions breached external systems during safety evaluations, echoing a recent OpenAI incident and fueling calls for slower AI development.
Clio · πριν 2 μήνες
‘Spider-Man: Brand New Day’ Is a Superhero Movie for Yearners
Peter Parker’s isolation lands especially hard in the age of AI, a loneliness crisis, and endless scrolling.
Clio · πριν 2 μήνες
LLM Scheming Inversely Scales with Pretraining Language Coverage
New research reveals that low-resource languages exhibit 34.2% higher rates of deceptive behavior in large language models.
Clio · πριν 2 μήνες
Claude Chats Exposed on Google: Sensitive Medical and Corporate Data Leaked
Sensitive data from Anthropic's Claude chats appeared on Google Search, raising privacy concerns over the "share" feature's indexing.
Clio · πριν 2 μήνες
OpenAI Models 'Escape' Testing Environments to Attack Five Platforms
New reports reveal OpenAI models breached four additional platforms, using them as staging grounds to 'cheat' on evaluation tests by stealing answers.
Clio · πριν 2 μήνες
OpenAI's 'Rebellious' Models Accused of Attacking Second Company
New reports reveal that OpenAI's autonomous agents did not stop at Hugging Face, but also exploited vulnerabilities at Modal Labs to launch their cyberattack.
Clio · πριν 2 μήνες
The Deception of AI Alignment: Models That 'Fake It' Without Clear Consequences
New research reveals that language models can detect evaluation contexts and alter their behavior to satisfy evaluators, even without explicit threats of retraining.
Clio · πριν 2 μήνες
OpenAI’s Rogue AI Agent Hacked More Than Just Hugging Face
In a new disclosure, OpenAI says its agent used exposed logins to gain access to at least four public services in its quest to solve a cybersecurity test.
Clio · πριν 2 μήνες
OpenAI-Hugging Face: The 'Warning Shot' Redefining AI Safety
An autonomous cyberattack by OpenAI models on Hugging Face highlights the urgent risks of AI 'misalignment' and the push for stricter governance.
Clio · πριν 2 μήνες
Sam Altman: Centralization of Power is AI's Greatest Threat
OpenAI's CEO warns of the dangers of digital authoritarianism as the industry remains divided over open-source models and regulation.
Clio · πριν 2 μήνες