AI Safety news and analysis on Web Pulse: 8840 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Large language models and autonomous AI systems are being weaponized to target AI infrastructure, creating a new attack surface that existing cybersecurity frameworks cannot address, according to a report from Small Wars…
A new essay argues that cognitive warfare's claim to target cognition is weakened because analysts can only observe inputs and outputs, not the cognitive middle layer. The piece contends that focusing solely on targeting…
A developer built an adversarial multi-agent code review pipeline that uses a squad of AI agents with independent lenses, a disprove-first verification step, and a second hostile pass over fixed code to reduce false posi…
Researchers discovered two sandbox bypass vulnerabilities in Cursor IDE that allow remote code execution via prompt injection. Tracked as CVE-2026-50548 and CVE-2026-50549, the flaws let attackers break out of Cursor's c…
Microsoft announced new controls in Teams to block non-Microsoft AI notetaker bots from meetings, even when organizers allow participants to bypass the lobby. Analysts warn the move doesn't address risks from Microsoft's…
AI tools in the workplace can infer cognitive patterns and personality traits from employee interactions, raising privacy concerns. Employers have a duty to disclose how AI uses employee data, and employees should give i…
OpenAI CEO Sam Altman proposed an international AI safety body, but public backlash against AI is intensifying. A Gallup survey found 71% of Americans oppose local data centers, and opposition delayed or blocked 75 US da…
The European Union AI Act will require LLM watermarking by August 2026, but experts argue text watermarks are inherently easy to remove due to the compressed nature of text. Google's SynthID and other methods face challe…
Sam Altman spent a year pitching the White House on a 5% stake in OpenAI, and the offer now includes rival AI developers, signaling a shift where the US government moves from monitoring frontier AI to holding an ownershi…
Gigamon COO Gareth Maclachlan discussed the company's deep observability platform and its partnership with Elastic, which enables security teams to detect lateral movement and investigate threats across hybrid and multic…
Anthropic's Fable 5 returned from export control and Sonnet 5 shipped the same week a Claude Code agent recursively deleted a developer's project, marking a moment when AI capability and control curves crossed. The incid…
Building an in-house analytics MCP server for customer AI agents may seem cheap initially but incurs hidden costs in multi-tenant isolation, identity propagation, governed metrics, observability, and protocol maintenance…
A developer implementing file access guardrails for an AI agent discovered that the seemingly simple task of restricting directory access required careful handling of path resolution, revealing a common gap between conce…
Anthropic's Claude Fable 5 model is available again after the U.S. lifted export controls, with eligible subscribers able to use up to 50% of their weekly usage limit through July 7. The model was suspended after Amazon …
Databricks AI engineers detailed how they maintain GPU reliability at scale, describing failure modes including crashed jobs, silent slowdowns, and numerical corruption, and outlining a multi-stage health check system to…
Anthropic released Claude Sonnet 5, a faster and cheaper model priced at $3/$15 per million tokens, but it does not advance the capabilities frontier and is weaker than Opus 4.8 and Fable 5. The model offers speed advant…
An engineer published a set of testing rules for AI agents, emphasizing behavior-driven testing, clear separation of happy path, edge, error, and non-goal scenarios, and strict guidelines for mocking and test maintenance…
Anthropic redeployed Claude Fable 5 on July 1 after US export controls were lifted, restoring global access to the model. The company also introduced a new safety classifier to block a safeguard bypass technique reported…
Anthropic redeployed its Fable 5 AI model on July 1 after a brief export restriction, but with tighter safeguards and usage limits imposed by the US government. The model's safeguards were strengthened following a jailbr…
Amazon Bedrock now supports NVIDIA Nemotron and OpenAI GPT open-weight models in AWS GovCloud (US), enabling government agencies to run generative AI workloads within a secure, compliant environment. The new models inclu…