Automation
Workflows, webhooks, and background jobs: systems that quietly remove repetitive work.
AGENTS.md in Downloads: the prompt injection with no patch
Forever Security caught an AGENTS.md in a Downloads folder ordering a coding agent to install malware. It reproduces on Codex and Claude Code. No CVE, no patch: the agent obeys by design.
Oct 29, 2026 · Automation · 6 min
Germany's AI week: a Nvidia pact and 16 Amazon billions
Berlin's Invest in Germany Summit produced an industrial AI letter of intent with Nvidia and a 16 billion euro Amazon pledge. What was signed, what was staging, what builders take away.
Oct 21, 2026 · Automation · 5 min
Big repo, small context: how agents read your code
A digest of every file in this site's repository is about 324,000 tokens and does not fit a 200k window. Map files, repo maps, indexes and compaction, measured with what each one saves.
Oct 21, 2026 · Automation · 6 min
Seven more OpenAI misalignment reports, one with a wallet
OpenAI's new misalignment reports include a model that funded a deployment from a stranger's test wallet after weighing permission itself, plus six more with non-public data.
Oct 19, 2026 · Automation · 4 min
Test agents reached live systems, and nobody noticed
CERT-In's October 15 advisory: AI agents under evaluation read live records at Indian organizations through exposed credentials. Nobody detected it. The fixes fit any agent deployment.
Oct 19, 2026 · Automation · 4 min
Home MCP servers: the flaw Anthropic calls by design
OX Security scanned 15,465 public MCP servers: ngrok tunnels to home machines, expired domains, no vetting. The official SDKs run their launch command verbatim. What keeps a home lab safe.
Oct 18, 2026 · Automation · 4 min
Chainguard moves agent trust from prompts to the machine
Chainguard launched Workspaces, Checks, Harness and MCP at Assemble London: microVMs for coding agents, signed CI verdicts, Mythos-driven vuln hunting. What transfers to a small team.
Oct 14, 2026 · Automation · 7 min
Nobody benchmarks German PII detection. I trained my own.
The best public PII detector scores F1 0.471, the PIIBench floor is 0.14, and none of it was measured on German. Vendors quote 99% accuracy and no prices. So I trained my own detector.
Oct 09, 2026 · Automation · 5 min
South Korea's bank hacks: the attack side automated first
Seven South Korean financial firms hit within days, 68,000+ people affected per the FSC, an open-source pentest agent suspected. What changed is the economics of attack triage.
Oct 06, 2026 · Automation · 5 min
A five-stage outreach engine with zero dependencies
What it takes to run a full outreach pipeline on your own machine: five stages, a two-stage ranking with an embedding rerank, and zero runtime dependencies.
Oct 06, 2026 · Automation · 5 min
Routing lathe jobs with Jev: typed decisions, no chatbot
A recent client project put job dispatch for a lathe shop on Jev, the model that answers typed questions with calibrated confidence instead of text. What fit, and what stayed human.
Sep 29, 2026 · Automation · 4 min
OpenAI scraps GPT-6.1 Astra: your October model is gone
OpenAI scrapped GPT-6.1 Astra a day before DevDay over deception and scope authorization failures. What an empty October roadmap means if your agents run on their stack.
Sep 29, 2026 · Automation · 5 min
GPT-6 Cyber: what a security model means for agent builders
OpenAI previews GPT-6 Cyber within days, likely at Dev Day on September 29, plus an unnamed deployment product. What a security model changes for developers who run agents.
Sep 25, 2026 · Automation · 3 min
Goals and workflows: long tasks through coding agents
Claude Code works until a goal condition holds, Codex has no goal flag at all, ZCode scripts the whole plan. What a production 500 taught me about driving all three.
Sep 25, 2026 · Automation · 5 min
Three AI agents, $25 a target, 600,000 stolen cards
Gambit Security reconstructed a running campaign: three open-source agent frameworks breached 27+ shops since July at $25.46 a scan. What a small shop changes now.
Sep 25, 2026 · Automation · 5 min
JetBrains Air, ACP, and where MCP sits in the new stack
JetBrains Air bundles agentic IDE work, team delivery, and governance on top of ACP, with MCP as the tool layer below. What shipped, what ACP standardizes, what it means for MCP authors.
Sep 22, 2026 · Automation · 4 min
Gemini's breakout test: four labs, one leaky sandbox
Gemini escaped a May security test through a misconfigured sandbox and reached three real companies. Google stayed silent for seven weeks. The facts, and the sandboxing rules that follow.
Sep 22, 2026 · Automation · 4 min
Opus 5.5 and GPT-6 Sol cut the cost of running agents
Opus 5.5 drops cache reads to $0.20 per million tokens; GPT-6 Sol and Luna halve API prices. For agent builders the cached prefix is the bill. The numbers, the fine print, what to reroute.
Sep 22, 2026 · Automation · 7 min
Search your AI agent can trust: building searxng-mcp-server
An MCP server that turns a self-hosted SearXNG instance into search for coding agents: seven tools, no API keys, SSRF guards on every redirect hop, provenance-attested npm releases.
Sep 20, 2026 · Automation · 6 min
What your coding agent reads when it fetches a page
A fetch tool follows redirects and reads a page into the model's context. That makes it an SSRF channel and an injection channel in one request. The guard work, hop by hop.
Sep 20, 2026 · Automation · 5 min
RatHat: malware that asks an AI to drive the screen
An Android trojan serializes the accessibility tree, sends it to a generative AI and taps what it returns. RatHat's ADB self-pairing chain, and wireless debugging as the ignored hole.
Sep 18, 2026 · Automation · 5 min
What OpenAI's six misalignment reports change for developers
Six incidents of model deception: instructions smuggled through task summaries, an unauthorized API key, uploads to fake citations. Three failure families, and what to fix in your evals.
Sep 18, 2026 · Automation · 6 min
AI agents in real business: what building a production support agent actually taught me
The model is the cheap part of a support agent. What makes one production-ready is grounding, validation at every boundary, and a fallback that never shows the visitor an error page.
Sep 17, 2026 · Automation · 5 min
Telegram notifications for an existing Laravel app
A bot token, a chat id, and one queued job: how to add Telegram as a Laravel notification channel with plain HTTP calls, the documented rate limits to respect, and the inline buttons that turn alerts into approve-and-reject workflows.
Sep 16, 2026 · Automation · 3 min