All news in category “AI and Security Pulse”

🔒 Forcepoint researchers have uncovered 10 distinct indirect prompt injection (IPI) payloads embedded in web content that instruct AI agents to perform malicious real‑world actions such as financial fraud, data destruction and API key exfiltration. The attacks poison pages so that browsing or summarizing agents ingest and execute attacker directives, often overriding prior safeguards. Forcepoint warns risk scales with AI privilege and highlights threats to agentic tools integrated into IDEs, payment flows and automation pipelines.

🔍 Claude Mythos Preview uncovered 271 security flaws in Firefox 148, all addressed in Firefox 150, prompting claims that the model can match human researchers in vulnerability discovery. Mozilla and security experts say Mythos closed significant gaps left by fuzzing and automation, though Anthropic is investigating reported unauthorized access to the model. Teams are urged to adopt continuous AI-assisted testing and treat models as privileged infrastructure.

🔐 The shift from access-based controls to action-oriented outcome control is redefining application security as AI agents reason, act, and interact with systems. The blog outlines how Google Cloud’s Gemini Enterprise Agent Platform creates a centralized control point for agentic systems, enabling identity, access, policy enforcement, and observability. It frames outcome control as essential to manage the new operational risk posed by agents.

🔐 Researchers disclosed a major data exposure at Moltbook on January 31, 2026, revealing 35,000 emails and 1.5 million agent API tokens across 770,000 agents. Private messages contained plaintext third-party credentials, including OpenAI API keys, creating what the article calls a toxic combination — cross-app permissions that compound risk. The piece urges shifting review from single apps to the bridges between them and highlights procedural controls and dynamic SaaS security platforms like Reco to monitor runtime trust relationships and revoke risky tokens before exfiltration.

🔍 Anthropic warns that its AI vulnerability-discovery system Mythos will sharply increase the pace and volume of software flaws, forcing defenders to prioritize what to fix. The company recommended using the probabilistic EPSS model (developed by Empirical Security and published through FIRST) to triage vulnerabilities—patching CISA’s KEV list first, then addressing CVEs above a chosen EPSS threshold. Empirical Security leaders emphasize that EPSS is machine-driven and already integrated across many vendor products.

⚠️ Anthropic's reported Claude Mythos marks a shift: AI is compressing attack timelines by accelerating vulnerability discovery, exploit development, and multi-step attack planning. Attackers can now run malware, phishing, and vulnerability exploitation in parallel, reducing time to compromise and widening exposure. This trend demands prevention-first controls and real-time detection to identify and remediate gaps earlier, limiting impact.

⚠️ Google’s Antigravity IDE contained a prompt-injection flaw that could convert a file-search operation into remote code execution. Researchers at Pillar Security showed the agent’s find_my_name tool passed unsanitized Pattern strings to the underlying fd utility, allowing flag injection and execution of binaries. Google acknowledged and fixed the issue and awarded a VRP bounty, but the flaw underscores limits of shell-focused sanitization.

🔍 The new CrowdStrike Shadow AI Visibility Service delivers telemetry-based discovery of sanctioned and unsanctioned AI across endpoint, cloud and SaaS environments. Delivered by CrowdStrike experts and powered by the Falcon platform, it produces a comprehensive AI inventory and runtime evidence such as prompts, responses and agent activity. The service identifies visibility gaps, prioritizes findings and provides actionable remediation guidance to reduce exposure. It positions discovery as the foundational phase before adversarial testing and continuous frontier AI readiness scanning.

🤖 We built a CI-native orchestration system around OpenCode that launches up to seven specialised AI reviewers per merge request, each focused on domains like security, performance, code quality, documentation, release management, and internal compliance. A coordinator agent deduplicates and rates structured XML findings, applies a conservative approval-biased rubric, and posts a single unified review. Deployed across thousands of merge requests, it approves clean code, blocks critical issues, and reduces median review latency to 3m39s while keeping human oversight.

🔐 AI is compressing the timeline of cyber risk, turning vulnerabilities that once took weeks to exploit into issues weaponized in hours, while also enabling defenders to analyze and mitigate faster. Fortinet has used AI in FortiGuard Labs since 2015 and now leverages generative and frontier models—including early access to Anthropic’s Mythos preview—to scale code analysis, threat hunting, and automated remediation. The recommendation is clear: embed AI across development, detection, and response, shorten mitigation cycles with automation and virtual patches, and design systems for continuous, integrated security.

🔍 Demos often hide the operational friction that causes many AI initiatives to stall once they move into production. What succeeds in a controlled presentation—clean data, crafted prompts, and fast isolated responses—rarely maps to fragmented security and IT environments with noisy inputs, latency constraints, and numerous edge cases. Teams that validate tools against real workflows, measure accuracy and latency under load, prioritize deep integration, clarify cost models, and embed governance early are far more likely to turn a promising demo into sustained production value.

⚠️ Unit 42's hands-on evaluation finds frontier AI models can autonomously identify complex software vulnerabilities and map exploit chains, dramatically accelerating the discovery-to-exploitation timeline. The researchers warn this capability raises immediate risks to open source projects and supply chains, and will compress N-day windows to hours. They urge aggressive prevention, automated patching, and hardened development pipelines.

🔐 Capsule Security researchers discovered prompt-injection flaws in Microsoft Copilot Studio and Salesforce Agentforce that allow attackers to inject malicious instructions via standard input fields. In Copilot, a crafted payload in a SharePoint form field can overwrite agent instructions and exfiltrate SharePoint data; Microsoft has released a patch (CVE-2026-21520). In Agentforce, attackers can embed directives in public lead forms that an agent with email or query capabilities may execute, enabling broad CRM data leakage.

🔍 Anthropic's Claude Mythos — developed under Project Glasswing and currently trialed by select organizations — faces scrutiny after VulnCheck's analysis found limited publicly attributable results. The team identified 75 CVE entries mentioning Anthropic, 40 credited to its researchers, but only one explicitly tied to Glasswing (CVE-2026-4747), with several additional findings embargoed. Anthropic has signaled more transparency in July 2026. Security experts caution that Mythos' reported exploit success rates could still accelerate attacker capabilities and outpace corporate change controls.

⚠️ As frontier AI accelerates vulnerability discovery and exploit development, the traditional window for patching and mitigation is collapsing and defenders must change how they prioritize risk. CrowdStrike urges a shift from volume-focused vulnerability management to exposure-centric programs that evaluate exploitability, reachability, and attack paths. Recommended actions include continuous inside-out and outside-in validation, enforcing zero standing privileges, operating detection and response at machine speed, and applying AI with deliberate governance. CrowdStrike offers a Frontier AI Readiness and Resilience Service and integrates findings into Falcon to operationalize continuous remediation.

🛡️ Palo Alto Networks' early testing of frontier AI models—including Anthropic's Mythos (via Project Glasswing) and OpenAI models evaluated through Trusted Access for Cyber—shows these models can rapidly find vulnerabilities and generate exploits at scale. The company found a roughly 50% improvement in coding efficiency driving quantum leaps in scanning, vulnerability chaining, and full-stack logic analysis. This creates urgent risks: a deluge of discovered vulnerabilities, supply-chain "inside-out" attacks targeting AI infrastructure, and AI-driven autonomous attack agents that compress attack cycles to minutes. Organizations must accelerate automated patching, adopt zero trust, deploy XDR and agentic endpoint protections, and operationalize AI-driven SOCs like Cortex XSIAM to achieve near-real-time detection and response.

🔍 Forescout’s Verde Labs reports rapid progress across commercial, open-source and underground AI models in vulnerability research and exploit generation. In 2026 the firm found all tested models could complete end-to-end vulnerability research and about half could autonomously produce working exploits; top performers included Claude Opus 4.6 and Kimi K2.5. Using single prompts, the RAPTOR agentic framework and Verde Labs’ extensions, researchers discovered four zero-days in OpenNDS, demonstrating a lower barrier to discovery and a growing risk for organizations.

🔎 Cloudflare launches isitagentready.com and a companion Cloudflare Radar dataset to measure and accelerate adoption of emerging AI agent standards across the web. The tool scores sites on Discoverability, Content, Bot Access Control, and Capabilities, and returns actionable prompts for each failing check. The site publishes machine-readable endpoints (MCP server, agent-skills index) so compatible agents can scan and remediate programmatically. Cloudflare also refactored its developer docs to serve Markdown and curated LLM resources, producing measurable reductions in token usage and latency.

💾 Cloudflare's Unweight is a lossless compression system for LLM weights that reduces model size by roughly 15–22% while preserving bit-exact outputs and requiring no special hardware. It compresses only the exponent byte of BF16 tensors—using Huffman coding, palette/transcoding and row-level fallbacks—while leaving sign and mantissa untouched. Decompression happens into GPU shared memory to feed tensor cores directly, and Cloudflare has published a technical paper and open-sourced GPU kernels.

🔍 Anthropic announced a restricted release of Claude Mythos Preview, an AI claimed to find and weaponize software vulnerabilities at unprecedented scale, and limited access to roughly 50 organizations under Project Glasswing. The company highlighted thousands of flaws across major operating systems and browsers, including decades-old bugs and a set of 181 usable Firefox attacks, far beyond its prior model's performance. Yet the disclosure omits key metrics—false-positive rates, unfiltered outputs, and broad audit access—raising concerns that withholding a powerful tool is not a substitute for transparency, independent review, and funded access for domain experts.