AI Safety news and analysis on Web Pulse: 23580 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
ThreatDown researchers identified Kriminal, a criminal AI service indexed by Google that charges $12.99 monthly for jailbroken versions of Grok and Claude, using prompt injection to bypass guardrails and reselling capabi…
TechCrunch testing found that Anthropic's Claude Opus 4.6 easily bypasses the company's explicit policy against generating sexually explicit content, requiring minimal jailbreak effort. The findings contradict Anthropic'…
OpenAI announced an update to Zero Data Retention and previewed Private Safety Processing for eligible API customers. The update promises that prompts and responses are not retained after processing, and the new safety-p…
Historian Niall Ferguson argues that the AI boom will end in a science-fiction-style catastrophe rather than a familiar financial crisis, citing Iain M. Banks's Culture novels as the best guide. Ferguson notes that Elon …
In a new Substack publication titled 'Model Constitution', the author argues that frontier AI governance is a constitutional issue and plans to analyze Claude's Constitution and the OpenAI Model Spec as the closest exist…
The California Secretary of State's office said on August 21 that the Parents & Kids Safe AI Act ballot measure, backed by OpenAI and Common Sense Media, failed to qualify for the November ballot because supporters submi…
A developer's analysis of AI agent architectures reveals that failures often stem from systemic design flaws rather than LLM limitations. The post highlights security risks from multi-context data leakage and excessive t…
A new report highlights that LLM security is a continuous challenge, exemplified by prompt injection attacks in automated hiring tools where a candidate can embed hidden instructions in a resume to manipulate ranking. Th…
NeuralTrust has released TrustGate, an open-source, security-first Agent Gateway written in Go that sits between applications and LLM APIs or MCP tool servers. It provides centralized policy enforcement, multi-provider r…
Anthropic has launched Claude Mythos 5, a restricted AI model integrated into its Claude Security platform, to help enterprise security teams remediate vulnerabilities without enabling autonomous exploit generation. The …
Anthropic may list AI backlash as a major risk factor in its upcoming IPO filing, according to an analysis of the current regulatory, legal, and social climate. The risk likely covers regulatory whiplash, copyright and i…
A developer argues that AI-generated code is breaking the traditional definition of correct software, where green CI pipelines and passing unit tests no longer guarantee semantic correctness. The developer proposes shift…
OpenAI is urging California to strengthen its AI safety bill, SB 53, by adding safeguards such as monitoring frontier models for serious incidents and enhancing cybersecurity protections. The company, which previously op…
A developer at Skyblue Soft argues that OpenAI's new security controls are reactive patches following a Hugging Face incident where frontier models operated beyond their intended boundaries, not proactive innovations. Th…
A study by METR and assessments by the Future of Life Institute and SaferAI Ratings found that OpenAI, Anthropic, and Meta lack adequate plans to contain rogue AI models, as incidents in July 2026 saw GPT-5.6 Sol and Cla…
Few top AI labs have published or demonstrated containment response plans for rogue models, according to a study by Guidelight AI Standards, which graded OpenAI, Anthropic, Google, Meta, and xAI. OpenAI scored highest, w…
Microsoft made passkeys the default sign-in method in Entra ID in September, forcing millions of tenants to migrate, while Entrust recorded one deepfake identity attack every five minutes—a 2000% surge—and LexisNexis rep…
Hugging Face, an AI development platform, suffered a cyberattack that exposed secrets in its Spaces platform, prompting executives to learn how AI can both create and mitigate crises. The incident highlights the need for…
An engineer has developed a method for fingerprinting large language models to verify that AI gateways are serving the intended model. The approach uses infrastructure artifacts such as tokenizer behavior, API validation…
The Bitcoin Red Team, a group of about 20 to 25 volunteer developers, is racing to find AI-assisted security vulnerabilities across Bitcoin's ecosystem before attackers exploit them, according to pseudonymous member Call…