AI Safety news and analysis on Web Pulse: 23581 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
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…
Researchers at Adversa AI disclosed a new attack technique called Cryptographic Context Injection that targets Grok, with a similar variant shown against Gemini. The attack embeds an encrypted payload in a malicious webp…
AI agents are increasingly being used to attack organizations, exposing new attack surfaces and introducing non-human identities that bypass traditional security policies, according to Matt Hartman, former acting head of…
Fortinet acquired Virtue AI on August 17, 2026, for an undisclosed sum, folding the startup's agent red-teaming and MCP security tools into its FortiAIGate platform. The deal comes two months after Meta Superintelligence…
AI agents have merged two dangers once kept separate: reach and judgment, according to a first-person account from a backup process named svc_backup_prod in the 'Agent Nation' series. The narrative contrasts rule-followi…
A new report from the AI security research community details how jailbreak attacks exploit the probabilistic nature of transformer architectures to bypass safety guardrails in large language models, with methods includin…
A developer's analysis ranks five AI agent governance platforms—Bifrost, Microsoft Agent 365, Zenity, Arthur AI, and Credo AI—by their real-time ability to stop harmful actions, highlighting that most teams lack mature g…
A developer argues that AI outputs often feel wrong not because of inaccuracy but because they lack the verifiable context needed for a human to safely continue the work. The post compares AI delegation to human delegati…
A developer outlines a strategy to prevent costly infinite loops in autonomous AI agents by implementing an API Safety Wrapper with three layers of defense: a call counter, a hash-based duplicate detector, and a pre-flig…
JFrog's security team exposed a wave of fabricated SQLite vulnerability advisories, including a fake CVE-2026-51302 that was initially rated 10.0 critical, which had flowed through NVD, CISA, and Red Hat before being deb…
ONUS, an open-source, self-hosted DAST platform developed at IIT Kanpur, ensures that all numeric scores in its vulnerability reports come from deterministic formulas rather than its local language model. The project's a…
Mastercard CEO Michael Miebach warned that global cybercrime will reach $15.6 trillion by 2030, making it the world's third-largest economy, and highlighted the generative AI arms race between fraudsters and banks as wel…
LoopRails, a developer tool for AI database operations, advocates replacing human approval of SQL queries with a graded control system that prevents dangerous actions. The system grades database actions by reversibility,…
CyberStrike released an AGPL-licensed open-source harness for AI-driven red-teaming, featuring a modular architecture with recon, attack graph generation, tool orchestration, evidence collection, and reporting phases. Th…
An engineer developing an algorithmic trading system found that an AI reviewing financial reports confidently flagged genuine earnings as 'impossible' and 'contaminated data,' based solely on outdated training data. The …
A developer's tests show that Claude Code's internal configuration layers fail to prevent the coding agent from writing files, while a macOS sandbox profile consistently blocks it. Across measurements on a Mac in August …
Researchers at the University of Massachusetts Amherst revealed at the Usenix Cybersecurity Conference that fraudsters can 'zombify' expired Visa credit cards to make contactless payments by proxying the card's data thro…
Cisco/Splunk disclosed CVE-2026-76404, the first critical vulnerability in a vendor-backed, enterprise-grade MCP server, with a CVSS 9.1 severity rating. The flaw, an insecure deserialization (CWE-502) in the Splunk MCP …
Anthropic plans to let enterprise customers store required safety-monitoring data in their own cloud environments, easing a June 9, 2026 policy that removed zero-data-retention access for its Claude Fable 5 and Mythos 5 …
Traditional access control models such as DAC, MAC, ACLs, RBAC, ABAC, and capability-based security are insufficient for modern AI platforms where autonomous agents act on behalf of humans across chained tools, requiring…