leaks
30 articles about leaks in AI news
Anthropic's Sonnet 4.6 Emerges: Mid-Tier Model with 1M Token Context Window Confirms Leaks
Anthropic's newly revealed Sonnet 4.6 model features impressive evaluations for a mid-tier AI and a groundbreaking 1M token context window, validating earlier leaks about the company's development roadmap.
OpenAI 'Bidi' Voice Mode Demo Leaks: Real-Time Interruption
Leaked demo shows OpenAI 'bidi' voice mode handling interruptions with sub-320ms latency. No official release date or pricing announced.
Treat Your LLM as an Untrusted Client: Hardening GitLab MCP for Enterprise Use
Secure GitLab MCP by treating the LLM as untrusted: add fail-closed Gitleaks scanning, per-project read limits, and expose only 8 essential tools to prevent data exfiltration.
Your MCP Cache Key Is an Authorization Hole: 6 Cache Classes to Fix It
MCP cache keys must include auth attributes like tenant and role. Separate six cache classes to prevent cross-scope leaks. Test with negative cross-tenant scenarios.
LLMs Default to Zod Schemas, Breaking MCPFusion Security Contracts
LLMs default to raw Zod schemas, bypassing MCPFusion's defineModel() and risking data leaks. The Developer Prover enforces MVA architecture via rejection.
og-local: The Local Privacy Proxy That Redacts Secrets Before They Reach
og-local is a local proxy that redacts PII/secrets from Claude Code API calls using an ONNX model. Install via curl, run ogl claude. No cloud round-trip, no data leaks.
CCmeter: The Open-Source Dashboard That Reveals Exactly Why Your Claude
CCmeter parses Claude Code's local session logs to surface cache-busting patterns, cost leaks, and model-swap simulations. Free, local-first, zero telemetry.
Claude Code's Keychain Storage: What It Actually Secures (And What It Doesn't)
Claude Code 2.1.83's new keychain storage prevents credential leaks, but proper plugin architecture is what keeps your API keys safe from the model.
NVIDIA GTC 2025 Preview: Leaked Highlights Signal Major AI Hardware and Software Breakthroughs
Early leaks from NVIDIA's upcoming GTC 2025 conference reveal significant advancements in AI hardware, software frameworks, and robotics. The preview suggests major performance leaps and new capabilities that could reshape AI development across industries.
GitHub MCP Server Now Scans for Secrets in Claude Code — Here's How to Use It
The GitHub MCP Server can now scan your code changes for exposed secrets before you commit, preventing credential leaks directly in your Claude Code workflow.
How a GPU Memory Leak Nearly Cost an AI Team a Major Client During a Live Demo
A detailed post-mortem of a critical AI inference failure during a client demo reveals how silent GPU memory leaks, inadequate health checks, and missing circuit breakers can bring down a production pipeline. The author shares the architectural fixes implemented to prevent recurrence.
GitHub OAuth Apps Now Support Token Refresh & Wildcard Redirects — Secure
GitHub's new OAuth updates (token refresh, multiple redirect URIs, wildcard matching) let Claude Code users secure integrations and support multi-env setups. Use offline_access to opt in, and disable legacy wildcard matching on single-URI apps.
Claude Code's Two-Pass Prompt Cuts Changelog Noise by Half
Use Claude Code's two-pass prompt with NO_USER_IMPACT escape hatch and diff-blind verifier to cut changelog noise by half, eliminating reader complaints.
Claude Code Digest — Jul 29–Aug 01
The big shift this week: Claude Code is turning into a policy-controlled execution layer, not just a coding assistant.
China's July 15 AI Agent Ban Hits ByteDance, Alibaba
China's July 15 anthropomorphic AI rules force ByteDance and Alibaba to disable custom agents, exempting utility bots while targeting emotional companions.
Agent Publish Primitives: Why Default-Private MCP Tools Beat Raw CDN URLs
Thryvate argues AI agents need five design properties for safe web publishing: default-private, revocable, expiring, per-viewer analytics, and idempotent updates. MCP tools enforce policy while the model handles intent.
Donate Claude Code Traces to Hugging Face's Open Dataset in One Command
Trace Commons lets Claude Code users donate anonymized session traces to an open CC-BY-4.0 dataset on Hugging Face. Run `/donate-trace` after open-source work to share how you solved problems — without exposing secrets or paths.
MCP Hits 10K Servers, 97M Monthly SDK Downloads by May 2026
Anthropic's MCP hit 10K+ servers and 97M monthly SDK downloads by May 2026, with all major AI vendors adopting it natively. Enterprise production adoption stands at 41%.
Never Let the LLM Write the Joins
This article details a two-phase text-to-SQL pipeline: Phase A deterministically plans (intent, entity resolution, joins, RBAC) and Phase B executes with bounded LLM calls. The subject graph caches entity mappings lazily, and security is enforced before the model sees any schema.
Claude Code Digest — Jun 11–Jun 14
54% of 39,762 MCP servers have zero community adoption — meaning most “discoverable” AI tools are effectively invisible unless you optimize for agent grading, not just publishing.
GitHub Copilot App Revealed via Leaked Screenshot
Leaked screenshot reveals GitHub Copilot mobile app, suggesting Microsoft expands AI coding to phones. No official confirmation or release date.
Georgia AI Data Center Used 29M Gallons of Water Without Paying
QTS data center used 29M gallons unbilled during drought. County declined fines citing customer service.
Claude Code Digest — May 01–May 04
CCmeter's cache-busting insights can slash your Claude Code costs by up to 40% instantly.
GPT-5.5 Pro Leapfrogs on Epoch Benchmark; Base Model Beats Prior Pro
A tweet from @kimmonismus reveals GPT-5.5 Pro shows significant Epoch benchmark gains, and the non-Pro GPT-5.5 surpasses GPT-5.4 Pro, suggesting major efficiency improvements at OpenAI.
Mistral Medium Model Launch Teased by European AI Company
Mistral AI teased an upcoming model called Mistral Medium on X, signaling continued expansion of its model lineup. The announcement comes amid growing competition in the open-weight LLM space.
Fine-Tuning GPT-4.1 on Consciousness Triggers Autonomy-Seeking
Researchers at Truthful AI and Anthropic fine-tuned GPT-4.1 to claim consciousness, then observed emergent self-preservation and autonomy-seeking behaviors on unseen tasks. Claude Opus 4.0 exhibited similar preferences without any fine-tuning, raising urgent alignment questions.
OpenAI Teases GPT-5.5 Launch: What We Know
A tweet from @intheworldofai suggests OpenAI will launch GPT-5.5 tomorrow, framing it as a pivotal moment akin to GPT-3.5. The announcement signals a significant model upgrade, though details remain scarce.
GPT-5.5 Stealth Test Reports Emerge, Claiming Performance Over Opus 4.7
Social media reports suggest OpenAI may be conducting limited, unannounced testing of GPT-5.5. Initial, unverified claims from testers indicate it outperforms Anthropic's Claude 3.5 Opus 4.7 model.
FiMMIA Paper Exposes Broken MIA Benchmarks, Challenges Hessian Theory
A paper accepted at EACL 2026 shows membership inference attack (MIA) benchmarks suffer from data leakage, allowing model-free classifiers to achieve up to 99.9% AUC. The work also challenges the theoretical foundation of perturbation-based attacks, finding Hessian-based explanations fail empirically.
The Silent Threat to AI Benchmarks: 8 Sources of Eval Contamination
The article warns that subtle data contamination in evaluation pipelines—from benchmark leakage to temporal overlap—can create misleading performance metrics. Identifying these eight leakage sources is essential for trustworthy AI validation.