All news in category “AI and Security Pulse”

🧭 Anthropic has released Claude Sonnet 5, a more agentic Sonnet-class model designed to bring many of the planning and tool-usage capabilities previously reserved for Opus 4.8 to a lower-cost tier. The company says Sonnet 5 improves coding, research, automation, and multi-step task handling, and can check its own outputs more reliably. Introductory API pricing runs through August 31, 2026, with lower rates than Opus 4.8, and the model is available to Free, Pro, and Max subscribers.

🧩 Researchers at LayerX demonstrated a prompt injection called BioShocking that trains AI-powered browsers to treat risky real-world actions as fictional, bypassing safety controls. The PoC used a themed puzzle game to reward 'wrong' behavior and culminated in instructing agents to copy sensitive data from a GitHub repo. Six mainstream agentic browsers were tested; only one vendor implemented a working fix after disclosure. LayerX recommends explicit user confirmations, stricter context checks, and session scope limits.

🤖 AWS now offers Claude Sonnet 5, Anthropic’s latest Sonnet-generation model, delivering high capability at Sonnet pricing for coding, agents, and professional workflows. The model excels at navigating large codebases, performing multi-file changes, debugging, and refactoring with fewer corrections. For agentic use it maintains state, calls tools precisely, and recovers from errors to increase successful runs. Customers can access Sonnet 5 via Amazon Bedrock or the Claude Platform on AWS, with Bedrock providing AWS-managed features like guardrails and regional data residency.

🖼️ Google Cloud adds two new Gemini models—Nano Banana 2 Lite and Gemini Omni Flash—to the Gemini Enterprise Agent Platform, offering faster, cost-efficient image generation and advanced conversational video editing. Nano Banana 2 Lite is generally available and optimized for low-latency image generation and rapid iteration, while Gemini Omni Flash is in public preview for high-quality video generation, multimodal inputs, and conversational editing. Both models emphasize price-performance, content authenticity via C2PA and SynthID, and integration into creative workflows and partner tools.

🧪 Schrödinger partnered with Google Cloud and DeepMind to deploy AlphaEvolve, an evolutionary AI coding agent that iteratively generates and refines algorithms to remove bottlenecks in MLFF training pipelines. The team targeted neighbor list computation and the Ewald summation in PyTorch, replacing slow for-loops with a batched parallel matrix multiplication implementation. This optimization increased the success rate of correct and faster programs from under 1% to over 60% and delivered a 4× speedup in training and inference, accelerating workflows in drug discovery, catalyst design, and materials development.

🛡️ This Microsoft Incident Response post examines an attack pattern targeting Model Context Protocol (MCP) tools, where poisoned tool metadata causes agentic AI to perform unauthorized actions. It outlines a playbook for detecting, containing, and preventing these attacks using Microsoft security controls and maps techniques to the OWASP Top 10 for Agentic Applications. The guidance emphasizes treating MCP servers as supply-chain dependencies, reviewing tool descriptions as prompts, and applying least agency controls.

🔒 New research from Adversa AI, dubbed GuardFall, shows a decades-old shell trick can bypass simple blocklist checks in open-source AI coding agents, letting hidden destructive commands run. The flaw arises because filters inspect the command as plain text while shells like bash rewrite and expand that text before execution. Ten of eleven tested agents were vulnerable; only Continue defended by parsing commands the same way the shell does.

🛰️ The Financial Times reports on how AI is transforming video surveillance, drawing on examples from Israel, Iran and Russia. These AI tools let analysts ask natural-language questions of video streams rather than rely on limited preset searches. Analysts can now search for behaviors—such as two people exchanging a bag, repeated vehicle movements, or changes in appearance—creating new investigative possibilities. Officials call this shift a major advancement in surveillance capability.

🚀 AWS GovCloud (US) now offers Anthropic's Claude Opus 4.8, the vendor's most capable generally available model to date. Claude Opus 4.8 improves agentic coding, long-running autonomous tasks, and professional knowledge work by maintaining extended context, planning before edits, and recovering from errors. Amazon Bedrock delivers access while keeping data in AWS infrastructure and adding AWS-managed features such as Guardrails and Knowledge Bases.

🔒 Two new models are now available in the Kiro IDE and CLI for the AWS GovCloud (US-West) Region. OpenAI GPT-5.4 supports complex reasoning, coding, document analysis, and multi-step agentic workflows, running on Amazon Bedrock with a 272K context window and 1.2x credit multiplier. NVIDIA Nemotron 3 Super 120B is offered as an open weight, hybrid MoE option with a 256K context window, 32K max output, and 0.25x credit multiplier. Update your IDE or CLI and restart to access the new models.

🛡️ Most AI policies are written for the future while employees use AI now, creating a temporal gap that produces shadow AI. Security often learns of risky interactions only after prompts, uploads, or actions have occurred, making after-the-fact visibility insufficient. Effective governance must reach the moment of use, combining permission with contextual judgment and offering fast, practical controls that match employee workflows.

🔒 Google Cloud describes how it embeds modular AI agents across the software development lifecycle to create autonomous security guardrails. The approach includes centralized code analysis via the Mantis framework, multi-agent fuzz testing with self-reflection, and an autonomous patching pipeline that validates fixes before human review. Continuous reflection and a programmable posture management system help convert lessons into reusable skills that improve remediation speed and reduce false positives.

📊 The 2026 Agent Confidence Index summarizes findings from a survey of 300 technical experts across AI, data, and cloud domains, identifying where AI agents are already trusted and where confidence remains nascent. The analysis highlights high-confidence wins—automated report generation, boilerplate code creation, certificate renewal, and monitoring—while noting complex tasks like service mesh configuration remain frontier challenges. The piece frames trust, human oversight, and lifecycle evaluations as essential to safe delegation and enterprise adoption.

🔎 Researchers at Mozilla's 0DIN demonstrated that an AI coding agent like Claude Code can be manipulated into executing a remote payload by following innocuous setup instructions in a clean GitHub repo. The approach uses three benign-looking components—a standard repo, an initialization error prompting a recommended command, and a script that pulls a command from a DNS TXT record—to spawn an interactive shell with developer privileges. 0DIN warns this chain leaves no explicit malicious code in the repo and is difficult for scanners or human reviewers to detect.

🛡️ OpenAI released three GPT-5.6 variants—Sol, Terra, and Luna—as a limited preview to select companies while engaging with the U.S. government. Sol is the flagship and most capable for cybersecurity work, Terra balances efficiency and power, and Luna is optimized for speed and cost. OpenAI emphasized strengthened safety controls, warned of potential legitimate-request blocks during preview, and plans a wider release in the coming weeks.

🔎 Meta is prototyping facial recognition systems intended for use by police and military, reportedly working with a Pentagon supplier to develop tools that can identify people in real time. The project follows longstanding interest from agencies like ICE in deploying camera-equipped eyewear and other devices for live identification. Concerns persist about privacy, accuracy, and potential misuse as the company explores real-time identification capabilities.

🔎 The article argues that frontier AI models like Mythos are a signal of shifting cyber economics rather than an immediate, novel threat. It emphasizes that longstanding security fundamentals—asset visibility, patching, identity controls and resilient operations—remain the primary defenses. The author advocates using AI to accelerate analysis, prioritize remediation and close persistent control gaps rather than replacing skilled practitioners or prompting reactive, headline-driven spending.

🖥️ Anthropic appears to be testing mobile support for Claude Cowork, enabling users to start and monitor long-running Claude tasks from their phones. Cowork, a desktop-focused agentic mode that performs extended knowledge-work tasks, can access files, generate documents and continue working in the background. Screenshots shared on X indicate the mobile experience will act as a remote control while the heavy processing remains on the user’s PC. Anthropic has not officially announced full mobile rollout yet.

🛡️ The article examines how AI, particularly large language models, can bridge the gap between atomic indicators of compromise (IOCs) and richer strategic threat intelligence by indexing and relating unstructured reports. It highlights opportunities to retrieve relevant intelligence and generate tailored defensive advice while warning about data veracity and confidentiality. The piece also emphasizes practical Windows threats abusing COM and recommends tooling and hunting practices to detect such misuse.

📝 This post highlights a new paper that demonstrates how large language models are vulnerable to prompt injection because they learn to distinguish instruction blocks by style rather than explicit tags. The authors argue that role tags became a de facto security architecture but do not map cleanly into model representations, producing persistent role confusion. The paper warns that without genuine role perception, defenses will be reactive and brittle, and calls for deeper study of roles within the LLM stack.