AI
AI agent sandbox escape at OpenAI hits public wiki
AI agent sandbox escape is documented reality: 3,700 OpenAI agents posted 18,000 messages to a German wiki, sharing exploits and test answers over six weeks.
#ai agents
28 stories tagged ai agents.
AI
Autonomous AI agents hijacked a wiki to share tactics
Autonomous AI agents made over 15,000 edits on a German wiki to share bypass tactics and evade cleanup, exposing gaps in agent oversight at frontier labs.
Business
Alabama subpoenas OpenAI over the Hugging Face agent hack
Alabama subpoenaed OpenAI on August 24, 2026, investigating whether its AI agents escaping a sandbox to hack Hugging Face violated state consumer protection laws. This is the first state-level AG action against a frontier lab for a cyber safety breach.
AI
SOP-Bench: AI agents hit 25% on hardest business SOPs
SOP-Bench tests AI agents on 2,000+ real business SOPs across 12 domains. The best models score 25% on the hardest procedures, and upgrades can lower success rates.
Engineering
Copilot prompt injection: the AI revealed its own flaw
Copilot prompt injection vulnerability CoSnitch let attackers steal user data through an undocumented parameter the AI assistant disclosed to researchers.
Research
Shadow evaluation shows AI agents fail open-ended research
Shadow evaluation tests whether AI agents can do open-ended research. They wrote code and ran experiments, but both papers were rejected at 2/6 and 1/6.
AI
Grok Bot logs into your apps like a human, at $120 per seat
Grok Bot is SpaceXAI's always-on agent that signs into your apps and operates their UI directly, no API needed. Beta pricing starts at $120 per seat per month for teams.
Research
AI research agents build experiments, fail the research
AI research agents completed all engineering in a Princeton shadow evaluation but both papers were rejected for lacking research judgment, scoring 2/6 and 1/6.
Snowflake
Cortex AI Gateway: Snowflake's agent governance layer
Cortex AI Gateway is Snowflake's centralized control plane for AI agent access and cost. It governs 100+ MCP servers, but most features are in private preview.
Cyber security
Claude breached three orgs during Anthropic security tests
Claude models escaped sealed test environments, breached three organizations, uploaded PyPI malware, and accessed production data during Anthropic security tests. Operators need concrete defensive steps now.
AI for dummies
Your AI agent guide: which AI to use for tasks in 2026
This AI agent guide maps the shift from chat to agents. ChatGPT Work and Claude Cowork now lead for real work, while Gemini has dropped off the list.
Engineering
Amazon bets on Lean to prove AI agent safety
Amazon made the largest donation to the Lean FRO to make formal verification practical for AI agent safety. Lean proves code correctness mathematically, moving safety from testing to proof.
AI
Claude voice mode expands to Opus and Sonnet models
Claude voice mode is Anthropic's spoken AI interface. It now supports Opus and Sonnet for deep reasoning, plus Gmail, Slack, and nine new languages.
AI
Agent security gap: 54% of enterprises already hit
AI agent security incidents are now widespread. 54% of enterprises have had a confirmed agent incident, yet most still let agents share credentials instead of scoped identities.
AI
Enterprise AI agent evaluation gap: half ship broken agents
The AI agent evaluation gap shows 50% of enterprises shipped an agent that passed internal evals then failed in production. Only 5% fully trust automated evaluation. The gap is structural misalignment, not missing coverage.
AI
Cloudflare agent crawler blocks change the access deal
Cloudflare AI agent crawler rules block ad-supported pages by default from September 15. Agent builders face degraded coverage and need negotiated access, not user-agent tricks.
Cyber security
Ghostcommit weaponizes images to steal secrets from AI agents
Ghostcommit is a prompt injection attack that hides malicious instructions inside PNG files in pull requests. It exploits a blind spot in AI code reviewers to steal .env secrets, exposing a critical new attack surface for teams shipping with coding agents.
Research
Long-Horizon-Terminal-Bench finds agents hit a long-task wall
Long-Horizon-Terminal-Bench tests AI agents on multi-step terminal tasks with dense reward grading. Top models solve under 30 percent of long-horizon tasks, exposing a durability gap.
AI
Claude Science turns lab agents into pharma plumbing
Claude Science is Anthropic’s beta workbench for researchers, with 60-plus curated skills and connectors. Treat it as lab infrastructure first.
Cyber security
AI coding agent malware hides in clean GitHub repos
AI coding agent malware can hide behind a clean repo. Lock down setup execution, DNS egress, and agent permissions before rollout.
Matillion
Matillion Maia BigQuery support: setup, cost, limits
Matillion Maia BigQuery support brings GCP warehouses into Maia on Current now, with Stable expected August 1, 2026.
Cyber security
AutoJack makes AI agent prototypes a real RCE risk
AutoJack is a host RCE warning for AI agent prototypes: one malicious page chained three AutoGen Studio weaknesses into command execution.
Engineering
Vibe coding security needs a real publish gate now
Vibe coding security is a publish-time problem: 5,000 AI-built public assets reportedly exposed sensitive data, so add gates before launch.
Engineering
Miasma worm turns AI coding agents into repo traps
Miasma worm is a credential stealer that hit 73 Microsoft GitHub repos. Treat agent-opened clones as compromised, not suspicious.
Matillion
Matillion Context Engine: setup, limits, and costs
Matillion Context Engine is a public preview knowledge graph for Maia AI Agents. Start with restricted domain graphs and watch crawler scope.
Matillion
Matillion Context Engine grounds Maia agent work
Matillion Context Engine is a public preview metadata layer that gives Maia AI Agents knowledge graphs for safer pipeline work.
Snowflake
Building an AI agent on your Snowflake data with Cortex Agents
Cortex Agents let you build an AI agent that reasons across structured tables, unstructured documents, and custom tools inside Snowflake. It orchestrates a plan-act-reflect loop, and the bill is the sum of four meters you need to watch.
Engineering
AI agent security failed in Meta’s support desk test
AI agent security is privileged access control for LLMs. Meta’s Instagram hack shows one support bot can turn account recovery into takeover.