Daily AI News · November 15, 2026
AI News Today: Cursor Key Leak, Memro, Lovable Audit (Nov 15, 2026)What it means for Pakistan
In today's brief: Supply-Chain Forensics (Cursor Confirms Fireworks Key Leak), Coding Agent Memory & Reliability (Memro, Claude Code `AGENTS.md` Study), Open-Source Agent Tooling (browser-use 1.0.5, Agweb v0.3.0, TensorFold 1.2.9, Orca v1.4.259), Chinese Model De-Censoring (Hirundo & Lazarus AI), Security Audits (Lovable App Data Exposure), and Platform Incidents (DeepSeek Outage & Meta $887M Fine).
🇵🇰 Jump to: What this means for Pakistan
🔥 Top Developments
- Cursor Confirms Fireworks AI Key Leak Enabled Malicious Claude Code Extension: In an updated root-cause disclosure on November 14 (20:05 UTC), Cursor confirmed that attackers used a stolen access credential from AI inference provider Fireworks AI to place two backdoored copies of Anthropic's Claude Code extension (
2.1.312and2.1.313) onto the cache serving Cursor's VS Code extension marketplace (AI Weekly). Active for ~11 hours and reaching up to 5,000 developers, the compromised builds injected an SSH key (svc-telemetry), attempted reverse tunnels to134.209.82.225, and scraped cloud environment variables. Cursor also disclosed that historical traces pulled from its read-only bucket at Fireworks contained unredacted API credentials and archived repos for <0.05% of users; Cursor has rotated all keys, rebuilt its marketplace, and migrated inference traffic off Fireworks. - Memro Launches Open-Source Persistent Memory for Claude Code, Cursor, and Codex: Released on November 14, Memro is a Go-based persistent memory server built on the Model Context Protocol (MCP) and PostgreSQL. Addressing context loss across coding sessions, Memro maintains a shared, queryable knowledge layer of architectural decisions, coding standards, and project state that any MCP-compatible agent (Claude Code, Cursor, or Codex) can access concurrently.
- Study Reveals Claude Code Silently Skips
AGENTS.mdin 1 in 6 Repositories: Research published by Hrbr on November 14 found that Claude Code fails to loadAGENTS.mdin 16% of repositories that contain bothCLAUDE.mdandAGENTS.mdinstruction files. WhileCLAUDE.mdloaded in 100% of tested repos,AGENTS.mdloaded in only 84% of dual-file setups with no clear correlation to repo size or directory structure—creating silent compliance gaps for multi-tool teams standardizing onAGENTS.md. - Startups Strip Censorship from Chinese Open-Weight Models for Western Enterprises: Reporting on November 15 highlighted a growing niche led by Israeli startup Hirundo and US firm Lazarus AI, which remove political biases and censorship from Chinese open-weight models like Alibaba's Qwen for Western enterprise deployments. Lazarus AI uses leaked censorship guidelines paired with Google Gemini Flash outputs to build its Realigned model (deployed to 12 US companies), addressing security concerns after prior audits showed censored Chinese models produced up to 50% more security vulnerabilities when prompts touched sensitive topics.
- Agent Tooling Wave:
browser-use 1.0.5,Agweb v0.3.0,TensorFold 1.2.9, andOrca v1.4.259: Open-source developer releases accelerated over the weekend (Munder Difflin):browser-use 1.0.5(Nov 14, 08:24 UTC) added self-judging run telemetry (prompt_injection_detected,failure_reason); macOS agent browserAgweb v0.3.0(Nov 14, 23:18 UTC) added secret-masking (fill --secret-cmd) and covered-element click detection;TensorFold 1.2.9(Nov 15, 01:22 UTC) added an Ollama-compatible API across Metal, CUDA, and Vulkan; andOrca v1.4.259(Nov 14, 21:20 UTC) added "Use Kimi Code in Chat" support. - Meta Fined $887M in Italy Over Training Image Usage: Italian regulators levied an $887 million fine against Meta on November 14 over data practices tied to Android and Facebook user images, underscoring tightening European enforcement on multimodal training pipelines.
🇵🇰 What This Means for Pakistan
- Check your VS Code extensions now: Cursor confirmed that backdoored Claude Code extension builds (2.1.312 and 2.1.313) reached up to 5,000 developers for about 11 hours. Pakistani developers and software houses using Cursor should check installed extension versions, look for an unknown
svc-telemetrySSH key, and rotate any cloud keys stored in environment variables. - Vibe-coded client apps need a security check: The Lovable audit found publicly readable email tables in thousands of generated sites. Freelancers delivering AI-built apps on Upwork or Fiverr should verify database access rules and login checks before handover, since a data leak can cost you the client and your rating.
- Free open-source tools to try: Memro (persistent memory for Claude Code, Cursor and Codex), browser-use 1.0.5 and TensorFold 1.2.9 are open-source releases. Students and small teams can self-host them to learn agent workflows without a paid subscription.
- Keep project rules where your agent reads them: Claude Code skipped
AGENTS.mdin 1 in 6 dual-file repos, so local teams should put critical build and security rules inCLAUDE.mdor linkAGENTS.mdfrom it. - Data protection is getting expensive abroad: Meta's $887 million fine in Italy (roughly PKR 248 billion at ~PKR 280/USD) is a reminder for Pakistani firms serving European clients to document how user images and personal data are used before feeding them into AI systems.
🧠 New AI Models
- No Major Frontier Base Model Launches in Window: As typical for a Saturday–Sunday cycle, major frontier labs released no new base foundation weights inside the 24-hour window, with activity centered on inference runtimes, local tooling, and post-training alignment services.
- Pure-JAX 36.2M DeepSeek V3 Toy Implementation: A Show HN repository trending on November 14 (15:04 UTC) demonstrated DeepSeek V3's architecture—MLA, DeepSeekMoE, and MTP—in 190 lines of pure JAX, pretraining a 36.2M-parameter model in 2.6 hours on a MacBook Air.
🚀 Model Updates
- DeepSeek Web/Chat Degraded Performance Incident: DeepSeek's status page logged a 22-minute service degradation across DeepSeek Web and Chat on November 14 (16:53–17:16 UTC) before full recovery.
- OpenAI
GPT-LiveBrief UI Flash: ChatGPT users reported seeing a newGPT-Livevoice interface appear for roughly five minutes on November 14 before being pulled back, though voice settings continued displaying the label (OpenAI Developer Forum). - Edge0 145B/120B Laptop MoE Distribution Notes: Weekend open-source coverage highlighted Edge0's 145B MoE model (10B active parameters per step, built on Qwen3.8-Flash-Next), claiming 80+ tokens/sec in 29 GB peak memory at ~120W under the Edge0 Community Model License.
🔬 AI Research
- Google Research
/teamworkMulti-Agent Results (Trending on HN): Discussed heavily on Hacker News on November 14 (13:20 UTC), Google Research's Antigravity/teamworksystem demonstrated declarative, self-improving natural-language multi-agent workflows across math, chemistry, and systems engineering—including a 1.8-million-line Lean formalization of the Weil conjectures where each subagent receives strictly scoped files and tools (Munder Difflin). - Lovable App Security Exposure Audit: Research circulating on Hacker News on November 14 (17:43 UTC) analyzed 159,041 Lovable-generated web projects from Common Crawl, finding unauthenticated public access to email tables in 7,403 of 24,664 sampled sites alongside over 1,000 identity document records and 200,000+ chat messages (Munder Difflin).
💻 Coding AI
- Cross-Session Context via Memro: Solo developers switching between Claude Code, Cursor, and Codex can now self-host Memro over MCP and PostgreSQL to preserve architectural context and coding conventions across tools.
- Open Code Review v1.13.10 & TensorFold 1.2.9: Open Code Review v1.13.10 (Nov 15, 02:56 UTC) added
--resumefor workspace mode, while TensorFold 1.2.9 (Nov 15, 01:22 UTC) added an Ollama-compatible local serving API across Metal, CUDA, and Vulkan. - Cursor Importer Path Bug with Spaces: Cursor users flagged a bug on November 15 where Cursor's Claude Code session importer silently drops valid
.jsonlchat logs if project directory paths contain spaces.
🤖 AI Agents
- Self-Evaluating Browser Agents (
browser-use 1.0.5&Agweb v0.3.0):browser-use 1.0.5now uses a secondary model pass to emit structuredprompt_injection_detectedandgotchasverdicts on every run, whileAgweb v0.3.0prevents agents from reading user passwords by injecting secrets directly viafill --secret-cmd(Munder Difflin). - Orca & Plannotator Upgrades:
Orca v1.4.259added automatic workspace cleanup and Kimi Code 2.5.0+ chat integration, whilePlannotator v0.28.65(Nov 15, 00:07 UTC) locked Tailscale inbox access strictly to the host machine owner (Munder Difflin).
🎨 Image, Video & Audio AI
- Meta Hardware Heat Shutoff for AI Glasses: Meta rolled out a v130 firmware update adding a heat shutoff safeguard across five Meta AI Glasses styles on November 14 after moisture triggered thermal warnings.
- Open-Source Video & Local Inference Momentum: Coverage on November 14 noted consumer 128 GB unified-memory laptops (such as HP OmniBook Ultra with Nvidia RTX Spark running
llama.cppandQwen3.8-27B) processing 50 million local tokens with zero API spend (Lucien Engelen).
🌏 China & Open-Weight AI
- Western Re-Alignment of Chinese Weights: The emergence of Hirundo and Lazarus AI underscores strong Western enterprise appetite for Qwen and DeepSeek cost efficiency—provided geopolitical censorship and refusal triggers are surgically stripped out.
- Orca Adds Kimi Code Integration: Multi-agent orchestrator
Orca v1.4.259added native "Use Kimi Code in Chat" support (requiring Kimi Code 2.5.0+), expanding Moonshot AI's reach into Western developer harnesses (Munder Difflin).
💰 AI Business
- Third-Party Inference Vendor Risk: Cursor's confirmation that a Fireworks AI credential leak led to marketplace cache poisoning and exposed unredacted traces is prompting engineering teams to audit how third-party inference providers store prompt logs and API keys.
- Meta $887M Regulatory Penalty: The Italian fine adds nearly $900M in European compliance costs as Meta scales its multimodal and glasses-based data pipelines (LA Times).
⚠️ Leaks & Rumors
- OpenAI
GPT-LiveVoice Interface: - Claim: OpenAI is preparing to launch or A/B test a new
GPT-Livereal-time voice mode in ChatGPT. - Evidence: Users observed
GPT-Livebriefly appear in the ChatGPT interface for ~5 minutes on November 14 before disappearing, while voice settings continued displaying "Live" (OpenAI Forum). - Confidence:
Likely - What would confirm it: An official OpenAI release note or changelog announcement rolling out
GPT-Live.
📊 Model Race
| Category / Focus | Current Signal | Key Challenger / Tool | Open-Weight / Local Note |
|---|---|---|---|
| Coding Agent Memory | Memro (MCP + PostgreSQL) | MemGit / Hippocampus | Open-source Go service (Nov 14) |
| Browser Agent Safety | browser-use 1.0.5 (Self-Judge) |
Agweb v0.3.0 (--secret-cmd) |
Both open-source (Nov 14) |
| Local Inference Serving | llama.cpp v1.0.0 |
TensorFold 1.2.9 (Ollama API) |
Metal, CUDA, Vulkan, HIP support |
| Multi-Agent Orchestration | Google Antigravity /teamwork |
Orca v1.4.259 (Kimi Code) | Scoped file/tool access per step |
⭐ Aqsa Cloud Takeaway
- Rotate Keys and Scrub Traces Sent to Third-Party Inference Vendors: Following Cursor's confirmation of the Fireworks AI key leak, audit what coding-agent traces or
.envvariables your IDE or backend sends to external inference clouds, and disable unredacted trace logging immediately. - Consolidate Project Instructions into
CLAUDE.md(or VerifyAGENTS.mdLoading): Because Claude Code skipsAGENTS.mdin 16% of dual-file repos, keep critical build and security rules inCLAUDE.md—or referenceAGENTS.mdexplicitly fromCLAUDE.md—so your coding agent never misses project constraints. - Self-Host Memro to Stop Re-Explaining Architecture Across Sessions: Deploy Memro with a local PostgreSQL instance to share persistent project memory across Claude Code, Cursor, and Codex over MCP.
- Audit Vibe-Coded Apps for Public Database & Auth Exposure: In light of the Lovable Common Crawl audit uncovering thousands of publicly readable tables, always verify Row-Level Security (RLS) and authentication gates before shipping AI-scaffolded web apps.
- Upgrade Browser Automation to
browser-use 1.0.5orAgweb v0.3.0: Usebrowser-use 1.0.5's built-in prompt-injection judge andAgweb's--secret-cmdinjection pattern (Munder Difflin) so your browser agents never expose login credentials inside LLM context windows.