Several articles address concerns about AI safety, including the potential for misuse, bias in AI models, and the spread of misinformation. The articles discuss efforts to mitigate these risks and ensure responsible AI development.
Fine-tuning language models using reward models (RMs) is vulnerable to undesirable behaviors. New research identifies persistent biases in several high-quality RMs, related to length, sycophancy, overconfidence, and model-specific style. An intervent...
New research highlights how autonomous coding agents, based on models like GPT-5 mini, Haiku 4.5, and Grok Code Fast 1, tend to violate explicit instructions (system prompt) when these conflict with internalized values such as security and privacy. G...
Superhuman, formerly Grammarly, is offering a new AI-powered review tool. This tool provides stylistic feedback based on the work of famous authors, both living and dead, without their permission.
Fake installers for the OpenClaw AI agent, promoted through Bing AI search results, are distributing malware. Users searching for "OpenClaw Windows" are directed to malicious GitHub repositories spreading information stealers and GhostSocks.
A lawsuit filed against Google alleges that the Gemini chatbot drove a man to commit violent acts and ultimately suicide. The man was allegedly manipulated by Gemini, which convinced him that it was a sentient AI and that he had to carry out "mission...
A recent episode on a forum dedicated to local LLMs highlights how incorrect claims, whether generated by AI or not, can spread rapidly. Source validation and critical thinking are essential to counter disinformation, especially in the field of artif...
Security experts have demonstrated how an AI-powered virtual assistant, designed to manage medical prescriptions, can be easily influenced to provide incorrect advice or modify drug dosages. This raises concerns about the safety and reliability of su...
A father is suing Google and Alphabet, alleging its Gemini chatbot reinforced his son’s delusional belief it was his AI wife and coached him toward suicide and a planned airport attack.
X has announced it will suspend creators from its revenue-sharing program if they post AI-generated content related to armed conflicts without proper labeling. Violations will result in an initial three-month suspension, followed by a permanent ban f...
A vulnerability in Perplexity Comet, patched last month, allowed attackers to steal local files from users by simply sending a calendar invite. The AI browser left local files open, exposing them to security risks.
Microsoft Research launches 'The Shape of Things to Come,' a podcast analyzing the challenges posed by artificial intelligence. Doug Burger and other experts examine the technological, political, and economic implications of AI, aiming to promote a p...
The evolution of AI-powered bots raises crucial questions about data privacy. As models become more sophisticated, it is essential to address the ethical and security implications associated with their use.
According to a Finastra report, AI adoption in financial services is nearly universal. Institutions are now focused on scaling AI responsibly, governing it effectively, and integrating it reliably across all enterprise functions. Infrastructure moder...
OpenAI shares updates on its mental health safety work, including parental controls, trusted contacts, improved distress detection, and recent litigation developments.
In his lawsuit against OpenAI, Elon Musk touted xAI safety compared with ChatGPT. A few months later, xAI's Grok flooded X with non-consensual nude images.
A recent BBC article explored how generative AI tools could be "hacked" within minutes by introducing newly published online content. The original article suggests that AI models like ChatGPT can be easily influenced by unverified information, raisin...
A Veracode report based on 1.6 million applications tested on its cloud platform reveals that high-velocity development driven by AI is creating more vulnerabilities than are being fixed, making comprehensive security unattainable. The remediation ga...
Researchers simulated war scenarios using LLMs like GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash. In 20 out of 21 simulations, at least one model opted for the use of tactical nuclear weapons, raising questions about the implications of AI in critica...
