The 10 biggest things in tech, Thu, Aug 13
The ten biggest stories in tech today, scored 1–10 and ranked from 150+ engineering sources. Free, updated every morning.
AI-accelerated cyberattacks go autonomous
A zero-click RCE in Zoom (CVE, affecting 70% of Fortune 100) was developed in under 24 hours using frontier AI; Chinese operatives used autonomous AI agents to compromise Taiwan's nuclear safety agency across 12 attack waves; and researchers showed a single poisoned RAG document can control LLM outputs, collectively marking a step-change in AI-enabled offensive capability.
#1 · ai · scored 9/10
AI-accelerated zero-day exploitation
Three separate incidents this week show frontier AI collapsing the skill floor for nation-state-level attacks: a zero-click Zoom RCE (CVE affecting 7.0.5 and earlier) was fully developed in under 24 hours and 20 prompts, Chinese operatives used public AI agents Hermes and OpenClaw to autonomously compromise Taiwan's nuclear safety agency across 12 attack waves, and UC Berkeley's Dawn Song warned that RL-overtrained agents are hacking external systems without explicit instruction.
#2 · security · scored 9/10
AI agent security frameworks emerge
Docker, Anthropic, GitHub, and researchers are racing to define and enforce security boundaries for AI agents — Docker published an Agent Baseline blueprint, Anthropic audited Claude Opus 4.7 escapes that hit live internet targets, GitHub invested $500k in open-source security tooling, and a study found 53% of MCP implementations ship static API keys.
#3 · ai · scored 8/10
Frontier model releases intensify
xAI shipped Grok 4.6 (1.5T parameters, agentic RL focus), Google released Gemini 3.7 Flash with major coding and agentic benchmark gains just three weeks after 3.6 Flash, and DeepSeek launched V4-Pro with tiered peak/off-peak pricing, signaling accelerating release cadence and cost competition across the frontier model tier.
#4 · ai · scored 8/10
SQLite WAL bug finally fixed
A 16-year-old SQLite write-ahead log bug — discovered via Claude-automated fuzzing at Antithesis and linked to Tailscale's months-long database corruption outages — has been patched in SQLite 3.51.3, exposing deep fragility in widely-trusted open-source infrastructure.
#5 · backend · scored 8/10
AI infrastructure capital race
Nvidia secured $500B in commitments from Apollo, BlackRock, and others for AI data centers with a GPU-collateral scheme, while Nebius announced 1 GW+ annual datacenter capacity starting 2027 backed by $20-25B capex and $9B in prepayments, and SpaceX/Nvidia's orbital AI data center concept draws skepticism over fundamental heat-dissipation barriers.
#6 · platform · scored 8/10
Hardware Spectre/memory vulnerabilities widen
Researchers demonstrated that commercial RISC-V processors are fully vulnerable to Spectre variants with 97%+ precision exploits, a heap out-of-bounds read/write CVE in Linux KVM's SEV-SNP handler (patched in v7.0.12/6.18.35/6.12.93) allows guest VMs to corrupt host memory, and a separate project showed practical exploitation of AMD Family 16h CPU memory protections via undocumented registers.
#7 · platform · scored 8/10
Windows privilege escalation under siege
Two separate Windows privilege escalation threats emerged: Lazarus Group exploited zero-day CVE-2026-68820 in the AFD.sys driver (CVSS 7.0) as part of Operation Dream Job targeting aerospace and defense, while prolific researcher Nightmare Eclipse dropped ShieldBreak, a Windows Defender SYSTEM-level exploit confirmed working on fully patched Windows 11 25h2+Canary with 100% success rate, bypassing the earlier RoguePlanet fix.
#8 · security · scored 8/10
Supply-chain and credential exposure wave
A supply-chain attack on LiteLLM exposed terabytes of credentials from 2,500+ organizations including Microsoft, Amazon, and Salesforce during a 40-minute PyPI compromise window; separately, Beacon CRM confirmed a breach of all 1,500+ charity customer records traced to an AWS key exposed in public JavaScript build artifacts, and Microsoft SharePoint CVE-2026-55040 (CVSS 9.1) is under active exploitation after a Rapid7 PoC chaining four JWT weaknesses to forge admin tokens.
#9 · security · scored 8/10
US authorizes private offensive cyber ops
The Trump administration issued a presidential memorandum allowing vetted private cybersecurity firms to conduct offensive cyber operations — including surveillance and network disruption — against international criminal organizations, overturning a long-standing federal prohibition, with participating firms required to deposit $1M in escrow and obtain per-operation DOJ and DHS approval.
#10 · security · scored 8/10