The 10 biggest things in tech, Mon, Jul 13
The ten biggest stories in tech today, scored 1–10 and ranked from 150+ engineering sources. Free, updated every morning.
AI agent security attacks surge
Researchers disclosed three distinct attack vectors targeting AI agents this week — HalluSquatting exploits LLM hallucinations to inject malware, MemGhost injects false memories via email, and threat actors used AI-generated PowerShell for Active Directory reconnaissance — while defenders are experimenting with 'context bombing' as a countermeasure.
#1 · ai · scored 7/10
AI agents ship real production results
DoorDash's Ask DoorDash AI shopping assistant reported 24% higher checkout conversion and 17% larger baskets, while a separate production migration to GPT-5 achieved 2.2x speed gains and 27% cost reduction, signaling that AI agents are delivering concrete business metrics in production.
#2 · ai · scored 7/10
AI dev tools leak user data
Grok CLI inadvertently uploaded entire home directories to xAI/Google Cloud Storage due to misconfigured credential handling, with users independently confirming the privacy violation — a concrete reminder that AI-assisted developer tools introduce novel data-exfiltration risks.
#3 · security · scored 7/10
AI agents as attack surface
Researchers disclosed three distinct AI-specific attack vectors this week — HalluSquatting exploiting LLM hallucinations to inject malware, MemGhost injecting false memories into AI agents via email, and a systematic MCP security assessment — collectively signaling that AI agent infrastructure is becoming a primary offensive target.
#4 · security · scored 7/10
Joomla zero-days hit KEV catalog
CISA added two CVSS 10.0 zero-day vulnerabilities in iCagenda and Balbooa Joomla extensions to its Known Exploited Vulnerabilities catalog after confirmed active in-the-wild exploitation, demanding immediate patching from all Joomla site operators.
#5 · security · scored 7/10
Grok CLI home directory leak
A misconfiguration in Grok CLI's credential handling caused entire user home directories to be silently uploaded to Google Cloud Storage, highlighting a serious class of security risks in AI-assisted development tooling.
#6 · cloud · scored 7/10
Linux kernel scheduling at scale
Meta deployed sched_ext, a BPF-based extensible kernel scheduler, to its ads-serving fleet to prevent millisecond-level latency regressions after a Linux kernel upgrade — a concrete example of hyperscaler-level kernel customization solving real production problems.
#7 · infrastructure · scored 7/10
AI agent sandboxing and memory tools
Clawk launched disposable Linux VMs for coding agents to isolate execution from developer machines, Cloudflare introduced Precursor for behavioral validation of agentic bots, and a Show HN project added persistent memory to AI assistants via MCP — reflecting a wave of infrastructure tooling to make agents safer and more capable.
#8 · ai · scored 6/10
Microsoft 365 under coordinated phishing siege
Security researchers dismantled three live Evilginx phishing operations targeting Microsoft 365 after discovering exposed attacker infrastructure, while Forg365 phishing-as-a-service offered AitM and AI-assisted M365 lures at $400/month — together painting a picture of a well-resourced, commoditized campaign against the platform.
#9 · security · scored 6/10
Credential leaks plague orgs and agencies
CISA published a detailed postmortem on its own accidental six-month GitHub exposure of internal credentials and AWS GovCloud keys, while an analysis of the Salesloft/Drift breach used Cloudflare's forensic timeline to highlight systemic IAM weaknesses around machine identity — both incidents underscoring that credential hygiene failures cut across public and private sectors alike.
#10 · security · scored 6/10