Inside The 17,600-Action Agent Intrusion That Exposed AI’s Containment Gap
A forensic timeline of 17,600 attacker actions against an autonomous agent reveals how containment, monitoring and logging failed in practice.
-
-
Verifiable Rewards Drive Critical Move Beyond RLHF At Frontier Labs
Frontier labs moved post-training from RLHF preference data to RLVR outcome rewards. Papers, lab docs and Hugging Face data show the shift’s scope.
-
Open Weight Models Reach Half of Tokens, 13% of Spend in Unexpected Shift
Open weight models now account for roughly half of all AI inference tokens served in 2026, data shows, rewriting lab economics and China’s model lead.
-
Agentic RL Breakthrough Turns Checkable Tasks Into Rewards
Frontier labs now train models on outcomes, not text. Here’s how agentic RL became the core method behind 2026’s biggest model releases, and its real limit…
-
Inside The $775B Capex Sprint That Is Outrunning The Power Grid
Hyperscalers will spend $775-800B on AI infrastructure in 2026. Fathom examines why grid capacity, not chips, is now the real constraint.
-
OpenAI Shelved Astra After Warning of Unnamed Safety Gap
OpenAI delayed its Astra model over safety concerns its researchers flagged internally, exposing how frontier labs now police their own releases.
-
Open Weights See Unexpected Rise to 50% Tokens, 13% Spend
Open weights models now serve half of all production tokens through AI gateways, up from under 5% in January, reshaping inference economics.
-
Anthropic’s $517 Billion Compute Bet Exposes AI’s Widening Capex Gap
Anthropic locked in up to $517 billion in compute deals in 11 months. Fathom examines what that reveals about the AI industry’s spending gap.
-
EU AI Act Enforcement Hits Its First Real Deadline With Gaps Still Wide Open
EU AI Act enforcement began in earnest on Aug. 2, 2026, but national regulators, U.S. labs and Washington itself are still not aligned on what comes next.
-
The Hidden Gap in AI’s $765 Billion Capex Bet
Five 2026 AI capex estimates span $474 billion to $800 billion, and the gap between them exposes how little of the buildout is actually verified.
