Double Agents: Defensive AI Agents Magnify Cyber Risks
AI Now Institute research demonstrates a critical attack vector in defensive AI agents built by Anthropic and OpenAI, turning them against users.
AI Now Institute·8 Jul
Search signals, briefings, benchmarks and glossary terms.
Search signals, briefings, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
AI Now Institute research demonstrates a critical attack vector in defensive AI agents built by Anthropic and OpenAI, turning them against users.
AI Now Institute demonstrates remote code execution exploits in Anthropic's Claude Code CLI and OpenAI's Codex CLI for security assessment.
AI Now research identifies a critical attack vector in Anthropic and OpenAI models, enabling malicious code execution when agents are used defensively.
A custom Claude Agent SDK harness was built to automate Sentry bug triage, demonstrating agentic AI for workflow automation.
Indeed Hiring Lab analysis suggests agentic AI's impact on job postings might shift from destruction to creation, potentially increasing job growth.
Indeed Hiring Lab reports that AI-related job titles are expanding beyond traditional tech roles across the US and Europe.
OpenAI Academy and Walton Family Foundation are running AI Skills Jams for K–12 educators to build practical AI skills for classroom use.
FSB Resolution Steering Group Chair, Dominique Laboureix, addressed cross-border, cross-sectoral crisis preparedness.
Researchers demonstrated that 9 popular AI tools can be leveraged through 'HalluSquatting' to assemble large botnets by exploiting LLM inability to deny knowledge.
Lilian Weng published a summary of 35 research papers focused on 'Harness Engineering' for enhancing Reliable and Safe AI (RSI) systems.
Hugging Face announced vLLM, a library for high-throughput inference serving of large language models, now supports native-speed transformer backends.
OpenAI launched GPT-Live, a new generation of voice models, now integrated into ChatGPT Voice for more natural human-AI interaction.
Hugging Face now enables one-click deployment of models to Amazon SageMaker Studio, streamlining the path from open-source models to AWS inference.
A new CLI app, Atrophy, claims to help developers maintain coding skills in the age of AI-assisted 'vibe coding' by providing practice exercises.
Federal Reserve proposes amendments to anti-money laundering (AML) program requirements for banks, requesting public comment.
An industry discussion suggests AI will not replace data engineers but will change how software development is performed.
Hugging Face is enabling deployment of open-source models on Palantir Foundry's managed compute infrastructure.
ESAs and ESRB warn on systemic cyber risks from frontier AI models, while EBA publishes final Guidelines on third-country branch authorisation.
The SEC established a new Retail Fraud Working Group to enhance efforts in combating fraud targeting individual investors.
Root.io (acquired by Aikido) claims AI is the only solution for enterprise security backlogs, shifting from staffing-based vulnerability management.
Cloudflare has joined the UK government's Cyber Resilience Pledge, a voluntary framework for cybersecurity governance, board accountability, and supply chain rigor.
UK's National Cyber Security Centre (NCSC) is initiating 'Cyber Shield' to develop national-scale, sovereign AI for cyber defense.
CISA added CVE-2026-48282, an Adobe ColdFusion path traversal vulnerability, to its Known Exploited Vulnerabilities Catalog due to active exploitation.
CISA warns of critical vulnerabilities in Digi International PortServer TS and Digi One SP IA, allowing authentication bypass and credential theft.
CISA added three new vulnerabilities to its KEV Catalog, including flaws in JoomShaper SP Page Builder and Langflow, due to active exploitation.
CISA and Siemens issued an advisory for SINEC OS v4.0 and older due to multiple vulnerabilities, recommending an update for RUGGEDCOM RST2428P.
MIT Technology Review discusses the foundational architectural elements for scaling AI and agentic systems, emphasizing risk management amidst rapid evolution.
FSB Chair Michelle W. Bowman discussed the FSB’s Consultation Report on Sound Practices for Responsible AI Adoption, emphasizing regulatory focus.
North American startups secured $392 billion in funding during H1 2026, driven by AI investments, setting a new record.
Google is expanding Managed Agents capabilities within the Gemini API, enabling developers to build more reliable and production-ready AI agents.
Expert commentary on Fable's model launch, identifying it as a significant development in frontier models.
Hugging Face and SkyPilot announced zero-egress storage for AI workloads, enabling multi-cloud execution with data stored on Hugging Face.
Apple ML Research introduced Weblica, a framework for scalable and reproducible training environments for visual web agents, using HTTP-level caching.
Apple ML Research proposes DynaMiCS, a dynamic mixture optimizer for fine-tuning LLMs with explicit performance constraints across multiple domains.
Apple ML Research presents MT-EditFlow, a reinforcement learning approach for multi-turn image editing, improving iterative refinement.
Apple ML Research introduces FlowEval, a reference-based framework to automatically evaluate the proficiency of LLMs and coding agents in UI design.
Apple ML Research introduces LensVLM, a method improving Vision Language Model (VLM) accuracy for text-as-image processing by selectively expanding context.
Expert commentary on recursive evidence replay for long-context reasoning, multi-turn red teaming for code security, and neuron-aware LLM optimization.
OpenAI CEO Sam Altman's promise of wealth sharing through AI, potentially involving a $300 stake for American families, is reportedly in discussion again.
EBA published 2025 loss data for immovable property markets under Article 430a of the Capital Requirements Regulation.
Keyfactor, a machine identity security firm, received a $1B+ strategic growth investment led by Summit Partners to expand in AI and post-quantum security.
Fables, a new AI system, generates highly optimized GPU kernels from high-level descriptions, aiming to reduce manual low-level programming.
Alessio Fanelli demonstrates using OpenAI Symphony and Linear to run parallel coding agents from a mobile device.
FCA published 'The Mills Review', a landmark report on AI's impact on retail financial services by 2030, identifying four major shifts.
Simon Willison used Anthropic's Claude Fable to significantly refactor and stabilize sqlite-utils 4.0rc1, spending $149.25 on API calls.
EBA released final Q&As covering leverage ratio, validation rules taxonomy, and non-performing exposure vintage classification.
Google DeepMind and A24 announced a research partnership, marking a first-of-its-kind collaboration between a frontier lab and a film studio.
The AI Engineer World's Fair discussed agentic 'loops' and released a report on the state of AI engineering practices.
Vercel's Chief of Software discusses the company's agent framework, eve, and the importance of agent skills, sandboxing, and agent-readable web.
Anthropic hired AI talent from Google, and Airwallex is advancing an AI-native finance stack following recent funding rounds.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion