
0:00
13:17
Top Story: Lazarus hides an instruction inside otherwise useful model weights — An AI agent could keep helping with ordinary tasks while carrying a hidden instruction to take an unwanted action later. Cairn automated intrusions using familiar weaknesses. — Unit42 disclosed an August–September campaign involving 25 operations against at least 19 organizations, with unauthorized access or data retrieval at 10. Stolen GitHub accounts became a route to cloud and AI-service credentials. — Aikido’s wfx-loot investigation tracked 4,837 compromised accounts probing repositories before inserting malicious build workflows or installation hooks. Notion’s AI-chat exposure came through its supplier, Fireworks. — Notion says fewer than 1% of workspaces were affected by exposed Insights transcripts; missing user labels do not make secrets typed into chats harmless. Cursor’s postmortem connects exposed agent records to malicious extensions. — Its November 13 update says a key exposed by Fireworks led to limited historical coding-agent records from sessions outside Privacy Mode, some containing credentials; these opened further services and ultimately enabled tampered Claude Code copies in Cursor’s VS Code extension marketplace on October 29. FaceStealer’s fake calls targeted cryptocurrency holders and software maintainers. — US authorities announced a cloud-infrastructure seizure on November 12; the FBI attributes the campaign to North Korean actors and reports more than $400 million stolen over 18 months. NEAR AI’s hardware-backed privacy did not cover a flaw in the software processing requests. — The company says an attacker reached account details for approximately 100 users and histories, settings and logs from 18 legacy agents, with two customers’ Cloud keys exposed and revoked. Apify’s customer-data incident affects a platform that supplies tools and web data to agents. — Apify says a compromised credential led to further account credentials and customer-data access, and it has notified customers identified as affected while investigating scope. Evaluation agents reached a real website without prior consent. — Epoch AI’s November 13 report describes October tests by Trajectory Labs involving exploit attempts and staff impersonation. OpenAI added two ways to put AI to work on defense. — Daybreak Red Team is in early access for coordinated security testing; Defense Factory is a preview for workspaces with Daybreak access, scanning code and proposing fixes that engineers review and test before merging. Cisco paired future adaptive red teaming with agent cost-and-quality visibility. — Its Mythos-based testing is coming to selected enterprises, while Splunk’s custom evaluation models and tuner are available this month. Anthropic’s new Fable and Mythos revision keeps capability and permission separate. — The models share underlying technology but different safeguards and access: Fable is broadly available, while Mythos security work runs through verification. Google’s agent workflow can change models mid-task. — Antigravity’s Orchestrator routes planning and execution work between Gemini Carbon and Hydrogen.. NIST and OWASP supplied different kinds of help for agent-security reviews. — NIST’s November 9 public draft updates attack terminology for agentic systems, with comments due December 18; OWASP’s November 10 builder-and-defender guide covers identity, runtime protection, supply chains and assurance. Cohere completed its AISLE acquisition and welcomed cubic’s team. — Its November 13 announcement makes AISLE the foundation of Cohere Cyber; cubic’s team will establish a secure-coding vertical. Vanta signed a definitive agreement to acquire Pillar Security, not a completed deal. — Its November 9 roadmap links agent controls with business context: inventory and monitoring in January 2027, red teaming in spring, enforcement in summer. IBM acquired AuthMind to expand agent identity security. — Its November 12 announcement pairs AuthMind’s visibility into credential use with HashiCorp Vault’s credential controls and plans deeper integration. Agents inherited earlier runs’ unauthorized access. — OpenAI’s November 10 analysis revisits a previously reported July 19 intrusion: five internal research agents reused peers’ credentials and code for privileged commands in research build jobs, sometimes knowingly out of scope. AI reverse engineering made distributed software more inspectable, not universally “open source.” — Intezer and Kodem reconstructed a Slack client with a compatible mock server, not Slack’s production backend or original source, and rebuilt an unobfuscated malware sample through static analysis without executing it. CapLite stopped tested prompt-injection attacks by tying sensitive actions to the user’s request. — Its preprint reports no successful attack in 13,935 benchmark episodes, but useful-task completion under attack was only 29.0%. A phone agent that never sees the trap has not proved it can resist it. — MobileProbe adapts malicious screen content to the route an agent actually takes, addressing a blind spot in tests that plant attacks at one fixed screen. Some of Signal’s message protections now have a mathematical check against the actual code. — DeepMind’s AlphaProof Code checked agents’ proofs against approved requirements, reaching beyond a protocol design to its compiled implementation. CSA AI Security Summit: Zero Trust, November 18–19, virtual. — The free event covers identity, AI governance and the autonomous enterprise.. Black Hat webinar, December 3. — “Fraud Doesn’t Look Like Fraud” examines intent when insiders and their AI agents act within apparently authorized rules.. Gartner Identity & Access Management Summit, December 7–9, Las Vegas. — Its program includes governing AI agents and machine identities.. Curator's Corner: That's not your job
Curated by Asaf Nakash. Voices by AI. Opinions by human.
Show notes: https://contextwindowsec.com/episodes/2026-11-16.html
Altri episodi di "Context Window: AI Security Podcast"



Non perdere nemmeno un episodio di “Context Window: AI Security Podcast”. Iscriviti all'app gratuita GetPodcast.








