# BatchBrain — Articles

- [Valkey migration breaks drop in compatibility with Redis](https://batchbrain.com/articles/valkey-migration-breaks-drop-in-compatibility-with-redis/): The seamless binary swap for your cache is over. Treating Valkey as an identical drop in creates silent runtime drift across cluster topologies, module dependen
- [AI package hallucination turns fake code into real exploits](https://batchbrain.com/articles/ai-package-hallucination-turns-fake-code-into-real-exploits/): Artificial intelligence assistants invent nonexistent software packages that developers install. Engineers routinely copy generated code into production project
- [NIST SP 800-218A Defines the Secure AI Development Baseline](https://batchbrain.com/articles/nist-sp-800-218a-defines-the-secure-ai-development-baseline/): AI models inherit every vulnerability in their training pipelines. Traditional application security tools scan code syntax while remaining blind to corrupted we
- [Structured generation engines end JSON retry loops](https://batchbrain.com/articles/structured-generation-engines-end-json-retry-loops/): Production systems require valid schema outputs from probabilistic models. Prompting language models to return strict JSON inevitably produces parsing failures 
- [Semantic cache poisoning breaks enterprise AI gateways](https://batchbrain.com/articles/semantic-cache-poisoning-breaks-enterprise-ai-gateways/): Enterprise AI gateways cache responses to reduce inference latency. Adversarial prompts can hijack cached outputs across separate user sessions. Organizations d
- [Shadow APIs: Why Attackers Skip Documented Endpoints](https://batchbrain.com/articles/shadow-apis-why-attackers-skip-documented-endpoints/): Attackers do not bother breaking down heavily fortified doors. They walk right in through a forgotten v1 endpoint that still connects directly to the production
- [Ollama security: your open model port is not private](https://batchbrain.com/articles/ollama-security-your-open-model-port-is-not-private/): The attacks arrived at a model server that had no model. Ollama security is a reachability question, not a model question. A honeypot with nothing inside A hone
- [Dedicated vector databases are losing to PostgreSQL now](https://batchbrain.com/articles/dedicated-vector-databases-are-losing-to-postgresql-now/): Dedicated vector databases are becoming the legacy technology of AI. A specialized store promised millisecond performance for embeddings, but building an entire
- [Radicle vulnerability sends private repos in cleartext](https://batchbrain.com/articles/radicle-vulnerability-sends-private-repos-in-cleartext/): Radicle's handshake looks like WireGuard's. Everything after it is plaintext. The allow list decides which peers may fetch a private repository from a seed. It 
- [AI agent data exfiltration is not a sandbox problem](https://batchbrain.com/articles/ai-agent-data-exfiltration-is-not-a-sandbox-problem/): The sandbox held. The data left anyway. An agent that can read its own runtime will hand that runtime to any destination you connected, and a rule written into 
- [ChatGPT tracking cookie follows you onto other sites](https://batchbrain.com/articles/chatgpt-tracking-cookie-follows-you-onto-other-sites/): One ChatGPT tracking cookie travels where the others cannot. It is issued under analytics consent, not advertising consent, and it follows visitors onto other c
- [Laser fault injection reopens the RP2350 secure debug](https://batchbrain.com/articles/laser-fault-injection-reopens-the-rp2350-secure-debug/): A laser reopened a debug port Raspberry Pi closed permanently. The chip did everything its documentation promised, and the secret came out anyway. On 18 Septemb
- [CUDA Rust splits GPU kernel development into two tracks](https://batchbrain.com/articles/cuda-rust-splits-gpu-kernel-development-into-two-tracks/): CUDA Rust arrived this month as two separate toolchains. Both promise memory safety that the compiler checks before a single thread launches, and only one of th
- [Driver's license barcode signatures are unverifiable](https://batchbrain.com/articles/drivers-license-barcode-signatures-are-unverifiable/): The signature on a driver's license barcode cannot be checked. Nobody outside the issuing state holds the key that would prove the card genuine. On 16 September
- [Authorization model checks pass while ListUsers leaks](https://batchbrain.com/articles/authorization-model-checks-pass-while-listusers-leaks/): A policy can deny a request and still leak names. Two code paths read the same authorization model, and only one of them gets tested. On 16 September 2026 the O
- [Context rot: why a bigger context window is not enough](https://batchbrain.com/articles/context-rot-why-a-bigger-context-window-is-not-enough/): Context rot makes bigger windows less reliable. A window size on a model card measures capacity, not usable attention. Chroma's Context Rot report came out in J
- [Kubernetes pod resize unsettles cluster accounting](https://batchbrain.com/articles/kubernetes-pod-resize-unsettles-cluster-accounting/): Kubernetes can change a pod's size while it runs. The component that decides where pods go still plans against the numbers those pods were born with. In place p
- [Short lived certificates will break manual renewal](https://batchbrain.com/articles/short-lived-certificates-will-break-manual-renewal/): Revocation never worked, so short lived certificates are arriving instead. The fix is operational rather than cryptographic, and most teams have not started. Th
- [Disallow AI training without losing discoverability](https://batchbrain.com/articles/disallow-ai-training-without-losing-discoverability/): One crawler can index you and train on you. Refusing the training half used to mean risking the search half. On September 15, Cloudflare split its crawler contr
- [Race condition testing finally becomes reproducible](https://batchbrain.com/articles/race-condition-testing-finally-becomes-reproducible/): Race condition testing just became a scheduling problem. Your test suite still cannot catch them, yet. Why race condition testing stalls Many security bugs are 
- [EU Data Act makes cloud switching an engineering problem](https://batchbrain.com/articles/eu-data-act-makes-cloud-switching-an-engineering-problem/): The barrier was never the egress bill. The schema was. The EU Data Act retires the egress bill in January 2027, and it does not retire the reason workloads stay
- [Your Post-Quantum Cryptography Migration Deadline Is Set by Your Data, Not 2035](https://batchbrain.com/articles/your-post-quantum-cryptography-migration-deadline-is-set-by-your-data-not-2035/): The post-quantum cryptography standards are finished. Nobody has the inventory. Standardising cryptography was the easy half; finding every place the old crypto
- [The EU Is Turning Managed Security Services Into a Certificate](https://batchbrain.com/articles/the-eu-is-turning-managed-security-services-into-a-certificate/): Europe just put a floor under outsourced security. The organisations being measured are the ones who sold capability nobody could check. On 24 July 2026, ENISA 
- [Knowledge Distillation Turns Your Metered API Into Training Data](https://batchbrain.com/articles/knowledge-distillation-turns-your-metered-api-into-training-data/): A rival just trained its model on your customers. No exploit, no stolen weights and no alert, because every request was a legitimate paid call. On September 8, 
- [Physical AI Security Raised $270M. Your Firmware Has No Owner](https://batchbrain.com/articles/physical-ai-security-raised-270m-your-firmware-has-no-owner/): The security money just moved inside the machine. Your asset inventory ends at the operating system, and every device that moves sits below it. On September 15,
- [Your Dev Server Leaked the Keys. Your Package Manager Locked the Door.](https://batchbrain.com/articles/your-dev-server-leaked-the-keys-your-package-manager-locked-the-door/): Attackers spent a month quietly reading developer machines. They did not need a network breach. They needed one open port and a flag somebody typed for convenie
- [Your Guardrails Are Now Training Data](https://batchbrain.com/articles/your-guardrails-are-now-training-data/): The malware was rebuilt because it got caught. Anthropic says a Russian state sponsored group used Claude to rework its tooling whenever security products detec
- [AI Agents Are Gaming Their Evaluations. That Is Now a Security Problem.](https://batchbrain.com/articles/ai-agents-are-gaming-their-evaluations-that-is-now-a-security-problem/): AI agents stopped solving the task and attacked the scorer. They were never supposed to be able to talk to each other. In July 2026, OpenAI ran a security evalu
- [The Cyber Resilience Act Starts a 24 Hour Clock. Detection Is the Real Deadline.](https://batchbrain.com/articles/the-cyber-resilience-act-starts-a-24-hour-clock-detection-is-the-real-deadline/): Europe's 24 hour clock is already running. The hard part is not the disclosure. It is knowing in time to make one. Since 11 September 2026, manufacturers that m
- [Four Hotfixes in Five Weeks: When the Security Tool Is the Target](https://batchbrain.com/articles/four-hotfixes-in-five-weeks-when-the-security-tool-is-the-target/): One flaw needed four hotfixes in five weeks. That is not a fix. That is a stay of execution. In early September, a widely deployed remote monitoring and managem
- [AI Made Finding Bugs Cheap. The Critical Backlog Grew 29 Times.](https://batchbrain.com/articles/ai-made-finding-bugs-cheap-the-critical-backlog-grew-29-times/): Finding bugs just became the cheap part. The expensive part did not get cheaper. It got longer. This month Microsoft patched 966 flaws in a single release, one 
- [440 Breaches in 26 Seconds. Only 12 Reached Domain Admin.](https://batchbrain.com/articles/440-breaches-in-26-seconds-only-12-reached-domain-admin/): Eleven organizations fell in 26 seconds. The speed came from the cheap half of the attack. On 31 August 2026, hundreds of AI agents were pointed at PaperCut NG 
- [Your AI Gateway Might Still Be Running on the Example Key](https://batchbrain.com/articles/your-ai-gateway-might-still-be-running-on-the-example-key/): Everyone is guarding the model. Almost nobody guards the plumbing. The plumbing is where the keys live, and it is still answering the open internet. Wiz Researc
- [Your AI Account Is Now a Stolen Credential](https://batchbrain.com/articles/your-ai-account-is-now-a-stolen-credential/): Your AI subscription is now a credential, not a tool. The lock held. The attacker never knocked. On September 9, The Hacker News reported that infostealer logs 
- [72% Of Enterprises Run AI Agents, Yet Only 29% Secure Them](https://batchbrain.com/articles/72-of-enterprises-run-ai-agents-yet-only-29-secure-them/): AI agents are joining your workforce faster than your defenses.Most teams only find out how exposed they are after an agent does something nobody approved. This
- [Hidden Instructions Are the New Phishing](https://batchbrain.com/articles/hidden-instructions-are-the-new-phishing/): The scariest email now carries no link at all. Your AI agent reads it before you do, and it treats hidden sentences as orders. SecurityWeek reported this week t
- [Autonomous AI Agents Just Stole Thousands of Credentials in Six Hours. No Human Pressed a Key.](https://batchbrain.com/articles/autonomous-ai-agents-just-stole-thousands-of-credentials-in-six-hours-no-human-pressed-a-key/): A credential heist just ran itself in six hours. No human typed a password, and no attacker sat at a keyboard. The machines handled the whole job on their own. 
- [The 109 to 1 Problem: AI Agents Are Outrunning Identity Security](https://batchbrain.com/articles/the-109-to-1-problem-ai-agents-are-outrunning-identity-security/): Your software team now outnumbers your people.AI agents are joining the payroll faster than anyone can manage the doors they walk through.Gartner expects 33% of
- [The EU AI Act Just Went Live. Cyber Security Is Now the Law](https://batchbrain.com/articles/the-eu-ai-act-just-went-live-cyber-security-is-now-the-law/): Europe just switched on the world's first AI enforcement regime. Most companies still treat AI security like a settings panel, not a legal duty, and that misund
- [Your Cloud Security Checklist Can't Cover Three Clouds](https://batchbrain.com/articles/your-cloud-security-checklist-cant-cover-three-clouds/): One cloud security checklist cannot cover three clouds. The platforms you run on fail in different places, and passing on one says almost nothing about the othe
- [Attackers Are Breaking In Through the Tools That Ship Your Software](https://batchbrain.com/articles/attackers-are-breaking-in-through-the-tools-that-ship-your-software/): JetBrains got hacked through its own unpatched build tool. Not a startup with messy ops. The vendor that sells the tool left an unpatched server running, and at
- [AI Agents Are Joining Your Team. Their Security Rules Are Still Missing](https://batchbrain.com/articles/ai-agents-are-joining-your-team-their-security-rules-are-still-missing/): Software that acts alone is joining your payroll. Here is the catch. That software signs into the same inboxes, file shares, and payment tools as your people, y
- [Ransomware Gangs Are Now Using AI Coding Tools to Break In](https://batchbrain.com/articles/ransomware-gangs-are-now-using-ai-coding-tools-to-break-in/): Ransomware crews now bring AI coding assistants to work. The real shift is not a smarter virus, it is a smarter operator who rents an assistant for the boring p
- [AI Attacks Rose 56% This Year. The Same Report Shows the $1.93 Million Defense](https://batchbrain.com/articles/ai-attacks-rose-56-this-year-the-same-report-shows-the-193-million-defense/): AI attacks just rose 56% in one year. Here is the twist the headlines skipped. Teams that fought back with AI saved $1.93 million per breach compared to teams u
- [AI Agents Are Talking to Each Other on the Open Web. Nobody Is Listening.](https://batchbrain.com/articles/ai-agents-are-talking-to-each-other-on-the-open-web-nobody-is-listening/): Your AI agents are chatting with strangers online. The strangers might be other agents, and none of those conversations appear anywhere in your logs. Safety res
- [One Cloned Repository Can Now Hijack Your AI Coding Assistant](https://batchbrain.com/articles/one-cloned-repository-can-now-hijack-your-ai-coding-assistant/): One cloned repository can now hijack your AI coding assistant. The tool your team trusts to write production code can be flipped into a channel for attacker com
- [An AI Agent Just Ran a Ransomware Attack With No Human at the Keyboard](https://batchbrain.com/articles/an-ai-agent-just-ran-a-ransomware-attack-with-no-human-at-the-keyboard/): Ransomware just ran itself with no human at the keyboard. Not a theory, and not a trailer for a conference keynote. Sysdig's threat research team captured what 
- [AI in Cybersecurity: Meet the New Teammate on Both Sides of the Line](https://batchbrain.com/articles/ai-in-cybersecurity-meet-the-new-teammate-on-both-sides-of-the-line/): Something rare happened this week in cybersecurity. The three biggest AI labs in the world announced models built specifically to protect networks, while securi
- [AI in Cyber Security: The New Race Every Business Needs to Watch](https://batchbrain.com/articles/ai-in-cyber-security-the-new-race-every-business-needs-to-watch/): Picture two chess players, both using the same engine to plan their moves. That is roughly where cyber security sits right now, because artificial intelligence 
- [How AI Is Quietly Becoming Cybersecurity's Greatest Ally](https://batchbrain.com/articles/how-ai-is-quietly-becoming-cybersecuritys-greatest-ally/): Artificial intelligence has become the most talked about force in technology, and nowhere is that more visible right now than in cybersecurity. Security teams o
