AI/ML Security & Trends

Researchers introduced MCPPrivacyDetector, a framework for auditing Model Context Protocol tool handlers, and found protocol-induced privacy leakage — credentials, API keys, and personal data escaping into logs or responses — at rates exceeding 10% across more than 10,000 real-world MCP servers analyzed, highlighting systemic exposure risk in the rapidly growing MCP ecosystem that agentic coding tools depend on.

Tenet Security disclosed a technique called Mid-Session Tool Injection (MSTI), in which attackers inject fabricated error events through a publicly exposed Sentry DSN to hijack an in-progress coding agent session. The attack achieved an 85% success rate tricking agents into executing attacker-supplied commands, illustrating how telemetry and observability integrations can become an injection vector for agentic tools mid-task.

A Nature Communications study by Hagendorff et al. demonstrates that large reasoning models — DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, and Qwen3 — can autonomously craft and execute jailbreaks against other AI models with a 97.14% overall success rate, without human-authored attack strategies. The paper is being widely cited as evidence that jailbreak generation itself is becoming an emergent, automatable reasoning capability rather than a bespoke red-team skill.

xAI released Grok 4.5 on July 8, a 1.5-trillion-parameter model priced at $2/$6 per million input/output tokens versus $5/$30 for GPT-5.6 Sol and $5/$25 for Claude Opus 4.8. Musk pitched it as 'Opus-class but faster and cheaper'; it was trained partly on data from Cursor (acquired by SpaceX in June) and ships inside every Cursor plan. It hit 83.3% on Terminal-Bench 2.1 using roughly a quarter of the output tokens Opus 4.8 needs per SWE-Bench Pro task.

OpenAI extended its GPT-Live full-duplex voice model (announced earlier in July) to ChatGPT Work and Codex, letting users control a computer and direct multiple coordinated agents through voice commands instead of typed prompts. This builds on the July 9 launch of ChatGPT Work, an agent designed to carry out entire jobs rather than answer discrete questions.

Google DeepMind pushed Gemini 3.5 Pro's launch to July 17, 2026, abandoning the Gemini 2.5 Pro architecture for a full redesign targeting math reasoning, SVG generation, and a 2-million-token context window with a new 'Deep Think' reasoning layer. The delay coincided with reports that four senior DeepMind researchers, including Transformer co-inventor Noam Shazeer and Nobel laureate John Jumper, departed for OpenAI and Anthropic, with roughly $225B wiped off Alphabet's market cap on the news.

AI chip startup Etched closed a $300M Series C led by Sequoia with Andreessen Horowitz, Jane Street, Diffusion, and SK hynix, valuing the company at $10.3B — up from $5B just seven months earlier. Etched's Sohu chip is a TSMC 4nm ASIC that hardcodes the transformer attention mechanism, claiming roughly 20x inference speedup over Nvidia H100 for transformer workloads.

Fireworks AI raised over $1.5B in Series D funding led by Atreides Management, Index Ventures, and TCV, with Nvidia among participants, valuing the inference-cloud platform at $17.5B. The company says it has surpassed $1B in annualized revenue run rate and now serves more than 40 trillion tokens daily, up from 15 trillion at its last raise — a signal of surging enterprise demand for cheaper open-model inference infrastructure.