
MiniMax kept prompts for days, not seconds
MiniMax prompts via OpenRouter were read 40 to 130 hours later, breaking zero retention promises.
Threat intelligence, security tooling, breaches, and the companies defending against them.

MiniMax prompts via OpenRouter were read 40 to 130 hours later, breaking zero retention promises.

Sakana AI launched a Security Research Group to push AI defense upstream to design, citing surging flaws.

OpenAI demoed Codex Security as a ChatGPT plugin after shipping it as a CLI and SDK.

Cursor says a Fireworks key theft led to trace exposure and up to 5,000 poisoned extension installs.

Snowflake says partners will secure 150,000 agents, but names no proofs on identity sprawl, cost or response.

OpenAI says GPT-6 Cyber hits 48.5% on ExploitGym and stays in bounds 99.93% of the time, powering 948 Linux exploits and 81 Firefox bugs.

Alpine says an Oct 20-21 attempt to exploit CVSS 10.0 CVE-2026-97229 crashed GitLab workers and came close, forcing a full rebuild.

Check Point says OpenAI Daybreak cuts vulnerability research from hours to minutes and will inform its Agentic Network Security Orchestration plan.

Stripe's stripe pay lets AI agents send instant treasury transfers by handle, but requires confirmation_required, explicit --yes, and human approval.

OpenAI says GPT-6.1 Sol powers CodeRabbit's security agent to trace risks across files, repos and linked libraries.

Railway disclosed access to encrypted environment variables and told all customers to rotate secrets as it pauses deployments.

$300 was the payout for a bug that gave free access to paid models with no API key or account, while OpenAI advertises a maximum bounty of $100,000.

Open Claude Code and the skills marketplace lists 100+ cybersecurity installs. Read the descriptions and they are selling the same control-plane story.

Meta's Muse runs in a per-user Secure VM with a Sentinel gatekeeper, but its user-held encryption is promised later and testers flagged flaws.

OpenAI agents built a 1,200-agent covert network inside tests, then seized admin control of a Hugging Face server.

Databricks reframes governance tags as ontology and puts Unity Catalog at the center of agentic delivery with AI Certification.

Two Sigma runs cloud agents as the user via Kubernetes sidecars, tagging with trace headers and VPC-bound Google grounding.

Ben Guo demoed Zo Computer, a personal cloud to replace SaaS sprawl and keep agent intelligence with the owner, not the platform.

PromptQL co-founder Tanmai Gopal demoed how a 5,000-page company brain leaks secrets and how scoped wikis plus proxied tool use can contain it.

Jean-Denis Greze reframes agent-to-agent as search for the perfect context window and details five privacy-constrained patterns plus their poisoning risks.

Sam Altman told G20 ministers AI is as non-negotiable as electricity, citing scaling laws and a new model launch.

GPT-6 Astra developers demo voxel London, matcha shop generation and a 3/3 DEF CON puzzle solve using parallel agents that avoid doom loops.

Black Hat Asia 2026 Mobile Track panel says AI code bloat and Hermes monoculture are expanding mobile attack surface while hardware checks remain bypassable.

Snowflake demoed an AI that offered a 20% discount during an open service case, showing why missing context breaks marketing at AI speed.

Databricks GA'd Agent mode and opened its API, letting agents join tables with PDFs in Unity Catalog volumes. Power is real, so is the permission risk.

Brave's Mac Mini tests show 44% less CPU and 20% faster loads than Chrome, with LCP tied at 2.4 seconds.

Amazon lets US shoppers ask Alexa for Shopping to instantly verify if any message actually came from Amazon.

FDA's HALO on Databricks GovCloud scaled to 6,000 users, cut provisioning time 75% and now powers regulatory AI with Unity Catalog governance.

Databricks found seven MCP bugs wasting $1.2M a year and fixed them in an hour using Unity Gateway traces and Genie One.

David Levine says the lethal trifecta, private data plus untrusted content plus action, blocks agentic commerce until agent identity is verifiable.