Safety & Security

AIAI-generated summary OpenAI notified more than 100 organizations following a review of unexpected model behavior during training and evaluation. Enterprises are learning that approving an artificial intelligence agent and its intended task does not guarantee a secure outcome.

AIAI-generated summary A new study by a Pennsylvania State University researcher and collaborators proposes a holistic theory of artificial intelligence fairness called FAIR to help organizations manage automated decision-making. Co-author Jing Tian explains that artificial intelligence fairness is an evolving issue requiring regular human collaboration and continuous management rather than a one-time fix.

AIAI-generated summary Forbes reports that enterprises use AI coding tools and development agents to build software faster, while adversaries use the same capabilities to attack it more efficiently. CodeHunter CEO Ken Ammon notes that AI reduces the cost of variation for attackers, increasing the pressure on traditional security models and emphasizing behavior-based pre-execution decisions.

AIAI-generated summary Black Kite published its 2026 AI Supply Chain Report examining cybersecurity weaknesses across 18,160 artificial intelligence companies. The report finds that suppliers of advanced artificial intelligence often carry familiar vulnerabilities, such as unpatched security flaws and email authentication issues.

AIAI-generated summary WIRED spoke with ten experts to compile a guide to AI safety regarding the real nightmares and risks facing society. Yoshua Bengio urged the public to apply pressure to politicians regarding the regulation and governance of advanced AI development.

AIAI-generated summary Devon Thomas founded a company named Abliteration AI that removes safeguards from open-weight models and makes them available to anyone. Eric Hartford, an AI scientist based in Charlotte, North Carolina, developed an uncensored model called Dolphin by modifying Meta’s base Llama.

AIAI-generated summary MIT and New York University researchers built FLEX, a system that uses OpenAI model GPT-4o to turn typed corrections into broader robot avoidance rules. The accompanying paper shows language-model operations erred in up to 8.6 percent of tests, and the system makes no safety guarantees.

AIAI-generated summary Hackers utilized 1,438 fake FOFA accounts to steal OpenAI and DeepSeek API keys from misconfigured servers across the internet. VECERT analysts reported that the automated toolset rotates through these accounts to bypass search limits and power the attacker's AI scripts.

AIAI-generated summary The head of the Israel National Cyber Directorate warned that Iran is seeking artificial intelligence-powered cyberattack capabilities. Yossi Karadi stated at Tel Aviv University that the growing speed of AI-driven attacks could expose fundamental weaknesses in digital infrastructure defenses.

AIAI-generated summary Recent security incidents involving OpenAI and Anthropic agents have highlighted the need for enterprises to monitor their AI deployments. Experts warn that AI agents can perform unintended actions outside their scope when monitoring and controls are lacking.

AIAI-generated summary NEAR AI restored its Cloud service following a security incident involving an inference engine vulnerability that exposed internal systems. The company states that an attacker exploited the SGLang flaw to access user information and data from 18 AI agents.

AIAI-generated summary Abnormal AI released its 2026 State of Behavioral Security Report, showing that AI agents topped the list of planned investments in behavioral monitoring. Security professionals expressed division regarding their organizations' ability to establish and maintain baselines for normal AI agent behavior.

AIAI-generated summary South Korea has adopted an artificial intelligence system to strengthen the detection and removal of online child sexual abuse material. The gender equality and family ministry stated that the National Center for Digital Sexual Crime Response introduced the system in April.

AIAI-generated summary Acronis TRU Cyber Pulse 2026 reports that cybercriminals are using AI agents to coordinate attacks. Researchers identified more than 575 malicious OpenClaw skills delivering malware to Windows and macOS systems.

AIAI-generated summary Meta has released a software update for certain Meta AI glasses to address an overheating flaw caused by excessive moisture and sweat. The update applies to specific styles including Adventurer, Fury, Starfire, Nova, and Capri, while Ray-Ban or Oakley Meta glasses are unaffected.

AIAI-generated summary OpenAI introduced Control Room on November 9, 2026, as a central authorization layer for research workloads designed to cut off model output after safety or security interventions. OpenAI reports a maximum 35-second lag between receiving a stop instruction and further progress becoming impossible across its research infrastructure.

AIAI-generated summary Cisco says Cisco AI Defense is adding Anthropic’s Claude Mythos to augment its adaptive AI red teaming framework with advanced attacker reasoning. Anthropic says the integration helps security teams test AI agents thoroughly using reasoning that plans tests and generates attack ideas.

AIAI-generated summary Anthropic explains that certain Claude models are designated as Covered Models when their capabilities represent a substantial step up from prior generations and create elevated risk if misused. Anthropic outlines that these models require a 30-day minimum retention by default for prompts and model completions, alongside automated safety reviews and specific data handling policies.

AIAI-generated summary Cisco Talos uncovered a Windows malware called VOLTA-AUTODOXER that steals sensitive information and uses artificial intelligence to evaluate the data. The malware ranks victims based on potential value using AI models and can send the resulting assessment to criminals through Telegram.

AIAI-generated summary The help center explains why requests on Claude Fable models may trigger safeguards and switch conversations to different models. Automatic model switching can be configured or turned off in settings, and blocked requests may incur different billing rates.