Field notes from a machine learning engineer specialising in cybersecurity. Master of IT at USC Adelaide, finishing Aug 2027.https://vikrantsharma.info/blogen-auhttps://vikrantsharma.info/blog/llama-cpp-hits-v0-1-0https://vikrantsharma.info/blog/llama-cpp-hits-v0-1-0Three years after first commit, the project that made local LLMs possible ships a stable release.Mon, 17 Aug 2026 00:00:00 GMTllama-cppmachine-learningopen-sourceinferenceversioningVikranthttps://vikrantsharma.info/blog/llm-trained-on-elementary-reading-levelshttps://vikrantsharma.info/blog/llm-trained-on-elementary-reading-levelsResearchers built an LLM using exclusively elementary-level text. The results challenge assumptions about what makes a model fluent.Sun, 16 Aug 2026 00:00:00 GMTmachine-learningllmtraining-datanlpVikranthttps://vikrantsharma.info/blog/thoughtdag-editable-context-graphs-for-llmshttps://vikrantsharma.info/blog/thoughtdag-editable-context-graphs-for-llmsA new interface shows chat threads as directed acyclic graphs where you can rewrite nodes and re-run paths. Fixes the branching problem most chat UIs ignore.Sat, 15 Aug 2026 00:00:00 GMTllmuser-interfaceprompt-engineeringgraph-theoryVikranthttps://vikrantsharma.info/blog/verifying-llm-generated-gpu-kernelshttps://vikrantsharma.info/blog/verifying-llm-generated-gpu-kernelsA new tool checks if ChatGPT's GPU kernels are actually correct before you run them in production.Fri, 14 Aug 2026 00:00:00 GMTllmgpuformal-verificationcudaVikranthttps://vikrantsharma.info/blog/flock-safety-lpr-privacy-guardrailshttps://vikrantsharma.info/blog/flock-safety-lpr-privacy-guardrailsThe company that sells license plate readers to 5000 police departments just announced privacy controls. Timing raises questions.Thu, 13 Aug 2026 00:00:00 GMTprivacysurveillancemachine-learningpolicyVikranthttps://vikrantsharma.info/blog/sqlite-wal-bug-survived-16-yearshttps://vikrantsharma.info/blog/sqlite-wal-bug-survived-16-yearsTailscale traced database corruption back to a WAL-mode edge case that survived billions of deployments since 2010.Wed, 12 Aug 2026 00:00:00 GMTsqlitedatabasesdebuggingconcurrencyVikranthttps://vikrantsharma.info/blog/stealing-reasoning-traces-from-llm-apishttps://vikrantsharma.info/blog/stealing-reasoning-traces-from-llm-apisResearchers figured out how to steal the internal reasoning steps from proprietary models like GPT-4o and Claude through API timing attacks.Tue, 11 Aug 2026 00:00:00 GMTllm-securityside-channel-attacksmachine-learningadversarial-mlVikranthttps://vikrantsharma.info/blog/needle2-14mb-llm-runs-on-wearableshttps://vikrantsharma.info/blog/needle2-14mb-llm-runs-on-wearablesNeedle2 runs on phones, smartwatches, and Raspberry Pis. The entire model is smaller than a single photo.Mon, 10 Aug 2026 00:00:00 GMTmachine-learningedge-computingllmhardwareVikranthttps://vikrantsharma.info/blog/sqlite-time-machine-zero-dependencieshttps://vikrantsharma.info/blog/sqlite-time-machine-zero-dependenciesA developer built a tool that lets you rewind SQLite databases to any point in their history, no external dependencies required.Sun, 09 Aug 2026 00:00:00 GMTsqlitepythondatabasesdebuggingopen-sourceVikranthttps://vikrantsharma.info/blog/100-gbps-packet-generator-go-af-xdphttps://vikrantsharma.info/blog/100-gbps-packet-generator-go-af-xdpA network engineer pushed Go to saturate a 100 Gbps link using AF_XDP. The surprising bit is not the speed, it is that Go got there at all.Sat, 08 Aug 2026 00:00:00 GMTnetworkinggolinuxperformancepacket-processingVikranthttps://vikrantsharma.info/blog/estrogen-masculinizes-mouse-brainshttps://vikrantsharma.info/blog/estrogen-masculinizes-mouse-brainsA 2010 study found that estrogen, not testosterone, drives sex-specific brain development in newborn male mice. The testicular hormone converts to estrogen in the brain.Fri, 07 Aug 2026 00:00:00 GMTneurosciencehormonesresearchbiologyVikranthttps://vikrantsharma.info/blog/kvm-guest-escape-zapscapehttps://vikrantsharma.info/blog/kvm-guest-escape-zapscapeZapscape is a guest-to-host escape for KVM on x86. It lets a compromised VM run code on the host machine, which is the nightmare scenario for cloud providers.Fri, 07 Aug 2026 00:00:00 GMTkvmvirtualisationsecuritycveVikranthttps://vikrantsharma.info/blog/hobby-programmers-reject-llmshttps://vikrantsharma.info/blog/hobby-programmers-reject-llmsWhy learning communities ban code assistants while enterprises mandate them.Thu, 06 Aug 2026 00:00:00 GMTllmlearningtoolingcultureVikranthttps://vikrantsharma.info/blog/llm-generated-cve-reports-polluting-sqlitehttps://vikrantsharma.info/blog/llm-generated-cve-reports-polluting-sqliteSomeone is flooding the National Vulnerability Database with fake SQLite vulnerabilities written by language models, and it is breaking actual security work.Tue, 04 Aug 2026 00:00:00 GMTsecurityllmsqlitecvesupply-chainVikranthttps://vikrantsharma.info/blog/defer-in-typescript-compiler-forkhttps://vikrantsharma.info/blog/defer-in-typescript-compiler-forkA developer forked the TypeScript compiler to add Go-style defer statements. The implementation works, the maintenance burden is real.Mon, 03 Aug 2026 00:00:00 GMTtypescriptcompilerslanguage-designgoVikranthttps://vikrantsharma.info/blog/encrypted-money-pipelinehttps://vikrantsharma.info/blog/encrypted-money-pipelineSixteen parsers, a server that writes what it cannot read, and what a hostile data source teaches you about production data engineering.Sat, 01 Aug 2026 00:00:00 GMTdata-engineeringcryptographycloudflaretypescriptVikranthttps://vikrantsharma.info/blog/hybrid-intrusion-detectionhttps://vikrantsharma.info/blog/hybrid-intrusion-detectionWhy my intrusion detection capstone needed an Isolation Forest, a Random Forest and an autoencoder to reach 0.95 F1 on real attack traffic.Sat, 01 Aug 2026 00:00:00 GMTmachine-learningsecurityintrusion-detectionpythonVikranthttps://vikrantsharma.info/blog/training-llms-on-8gb-gpuhttps://vikrantsharma.info/blog/training-llms-on-8gb-gpuSomeone just released a minimal codebase for supervised fine-tuning, direct preference optimisation, and group relative policy optimisation on consumer hardware.Sat, 01 Aug 2026 00:00:00 GMTmachine-learningllmfine-tuninggpuopen-sourceVikranthttps://vikrantsharma.info/blog/llm-code-review-globaleaks-41-findingshttps://vikrantsharma.info/blog/llm-code-review-globaleaks-41-findingsA security firm ran Claude through GlobaLeaks' codebase and found medium-severity bugs for seventy-six dollars per finding. The question is whether a human would have caught the same issues faster.Fri, 31 Jul 2026 00:00:00 GMTllmsecuritycode-reviewappsecowaspVikranthttps://vikrantsharma.info/blog/llm-router-deprecated-manifesthttps://vikrantsharma.info/blog/llm-router-deprecated-manifestManifest deprecated their LLM router after six months. The reason is not what I expected.Fri, 31 Jul 2026 00:00:00 GMTllmperformancearchitecturelatencyVikranthttps://vikrantsharma.info/blog/llm-honeypot-catches-ai-scrapershttps://vikrantsharma.info/blog/llm-honeypot-catches-ai-scrapersA researcher deployed a fake human verification page that only AI crawlers would fall for. The logs are filling up.Wed, 29 Jul 2026 00:00:00 GMTllmweb-scrapingcybersecurityhoneypotVikranthttps://vikrantsharma.info/blog/openai-published-codex-security-playbookhttps://vikrantsharma.info/blog/openai-published-codex-security-playbookOpenAI open-sourced the internal security guidelines they used when building Codex. Turns out threat modelling an AI code generator is different from threat modelling a database.Wed, 29 Jul 2026 00:00:00 GMTsecurityllmcodexopenaithreat-modellingVikranthttps://vikrantsharma.info/blog/apple-patches-230-vulns-in-one-dayhttps://vikrantsharma.info/blog/apple-patches-230-vulns-in-one-dayiOS 26.6 fixed 75 security issues, macOS got 155. That is not normal patch volume.Tue, 28 Jul 2026 00:00:00 GMTvulnerability-managementapplepatch-tuesdaydetection-engineeringVikranthttps://vikrantsharma.info/blog/github-security-team-orchidfiles-breachhttps://vikrantsharma.info/blog/github-security-team-orchidfiles-breachOrchidFiles reported a flaw that let anyone read private repo secrets. GitHub marked it duplicate, did not patch it, then banned the researcher.Sun, 26 Jul 2026 00:00:00 GMTgithubsecuritybug-bountydisclosureVikranthttps://vikrantsharma.info/blog/what-llm-labs-actually-hire-forhttps://vikrantsharma.info/blog/what-llm-labs-actually-hire-forResearch engineer roles at the big labs filter for production ML skills first, paper count second.Sun, 26 Jul 2026 00:00:00 GMTmachine-learningcareerllmresearchVikranthttps://vikrantsharma.info/blog/debian-votes-to-ban-llm-contributionshttps://vikrantsharma.info/blog/debian-votes-to-ban-llm-contributionsA Debian general resolution proposes banning LLM-generated patches. The reasoning is direct: you cannot verify the training data's licensing.Fri, 24 Jul 2026 00:00:00 GMTopen-sourcellmlicensingdebianpolicyVikranthttps://vikrantsharma.info/blog/encryption-going-dark-debate-fifteen-years-laterhttps://vikrantsharma.info/blog/encryption-going-dark-debate-fifteen-years-laterLaw enforcement claimed strong encryption would end investigations. A new paper tracks what actually happened: encryption went mainstream, crime still gets solved.Fri, 24 Jul 2026 00:00:00 GMTencryptionsecurityprivacye2eepolicyVikranthttps://vikrantsharma.info/blog/kids-treat-chatbots-like-people-more-than-adults-dohttps://vikrantsharma.info/blog/kids-treat-chatbots-like-people-more-than-adults-doNew research shows children anthropomorphise LLMs at much higher rates than adults, which changes how we should think about guardrails.Thu, 23 Jul 2026 00:00:00 GMTai-safetyllmresearchethicsVikranthttps://vikrantsharma.info/blog/cisco-antares-small-models-for-securityhttps://vikrantsharma.info/blog/cisco-antares-small-models-for-securityAntares models are 1B to 8B parameters, fine-tuned on security tasks, and Apache 2.0 licensed. This is not another rebranded Llama wrapper.Wed, 22 Jul 2026 00:00:00 GMTmachine-learningcybersecurityllmthreat-detectionVikranthttps://vikrantsharma.info/blog/kimi-k3-fixes-security-bugs-codex-refusedhttps://vikrantsharma.info/blog/kimi-k3-fixes-security-bugs-codex-refusedA Chinese LLM patched critical security bugs in a codebase where OpenAI's models and Anthropic's Claude refused to engage. The refusal problem is real.Mon, 20 Jul 2026 00:00:00 GMTllmsecuritymodel-alignmentkimiVikranthttps://vikrantsharma.info/blog/open-source-ai-fight-same-as-softwarehttps://vikrantsharma.info/blog/open-source-ai-fight-same-as-softwareDavid Siegel argued with Richard Stallman for two years in the 1980s. Now he is watching the exact same debate play out with AI models.Mon, 20 Jul 2026 00:00:00 GMTopen-sourceairegulationhistoryVikranthttps://vikrantsharma.info/blog/qubes-os-security-public-recordhttps://vikrantsharma.info/blog/qubes-os-security-public-recordFifteen years of a compartmentalised desktop OS, zero remote code execution bugs. The paper tracking every public Qubes vulnerability is out.Sat, 18 Jul 2026 00:00:00 GMTqubes-ossecurityvirtualisationsupply-chaincveVikranthttps://vikrantsharma.info/blog/agent-security-is-a-systems-problemhttps://vikrantsharma.info/blog/agent-security-is-a-systems-problemA new paper argues that LLM agent security cannot be solved by better prompts or guardrails alone. The architecture itself leaks privilege.Fri, 17 Jul 2026 00:00:00 GMTllm-securityagent-systemsprivilege-escalationsandboxingVikranthttps://vikrantsharma.info/blog/sqlite-nul-character-stringshttps://vikrantsharma.info/blog/sqlite-nul-character-stringsSQLite will happily store NUL characters in text columns. Your application layer might truncate them silently.Wed, 15 Jul 2026 00:00:00 GMTsqlitedatabasesc-interopdata-validationVikranthttps://vikrantsharma.info/blog/llm-guardian-angels-personalised-assistantshttps://vikrantsharma.info/blog/llm-guardian-angels-personalised-assistantsGwern argues personalised LLM assistants could filter spam, draft replies, and catch your mistakes before you send them. The privacy trade-off is obvious.Tue, 14 Jul 2026 00:00:00 GMTllmsecurityprivacypersonalisationVikranthttps://vikrantsharma.info/blog/neural-network-in-sqlhttps://vikrantsharma.info/blog/neural-network-in-sqlA developer trained MNIST digit recognition using only SQL queries. No Python. No frameworks. Just recursive CTEs and window functions.Tue, 14 Jul 2026 00:00:00 GMTsqlneural-networksmachine-learninghacker-newsVikranthttps://vikrantsharma.info/blog/mcp-security-report-2026https://vikrantsharma.info/blog/mcp-security-report-2026Anthropic's Model Context Protocol promised to standardise how AI agents talk to tools. A new audit shows most implementations ship with auth disabled by default.Sun, 12 Jul 2026 00:00:00 GMTsecurityai-agentsmcpauthenticationllmVikranthttps://vikrantsharma.info/blog/mesh-llm-distributed-inference-on-irohhttps://vikrantsharma.info/blog/mesh-llm-distributed-inference-on-irohIroh built a system that splits LLM inference across volunteer nodes. The networking stack handles dropouts mid-inference. Wild.Sat, 11 Jul 2026 00:00:00 GMTllmdistributed-systemsp2pinferenceirohVikranthttps://vikrantsharma.info/blog/gemini-2-5-flash-discontinuation-panichttps://vikrantsharma.info/blog/gemini-2-5-flash-discontinuation-panicGoogle announced Gemini 2.5 Flash will be discontinued in February 2027. Developers who built on it are scrambling.Fri, 10 Jul 2026 00:00:00 GMTmachine-learningllmsapi-designproduction-mlVikranthttps://vikrantsharma.info/blog/llm-burnout-is-realhttps://vikrantsharma.info/blog/llm-burnout-is-realThe moment you realise you are debugging prompt chains instead of writing code.Fri, 10 Jul 2026 00:00:00 GMTllmmachine-learningengineeringburnoutVikranthttps://vikrantsharma.info/blog/llm-coding-variance-matters-more-than-benchmarkshttps://vikrantsharma.info/blog/llm-coding-variance-matters-more-than-benchmarksDan Luu measured the same coding task twenty times with the same prompt. The variance in output quality was higher than the difference between model versions.Wed, 08 Jul 2026 00:00:00 GMTllmcodingvariancebenchmarksclaudeVikranthttps://vikrantsharma.info/blog/polyml-standard-ml-still-compiling-after-40-yearshttps://vikrantsharma.info/blog/polyml-standard-ml-still-compiling-after-40-yearsStandard ML has a production compiler that predates Java, runs on ARM64, and ships with a garbage collector tuned for symbolic computation.Tue, 07 Jul 2026 00:00:00 GMTstandard-mlcompilersgarbage-collectionfunctional-programmingVikranthttps://vikrantsharma.info/blog/sqlite-utils-schema-migrationshttps://vikrantsharma.info/blog/sqlite-utils-schema-migrationsSimon Willison's sqlite-utils 4.0 adds schema migration tracking. The library that made it trivial to load data into SQLite now makes it trivial to evolve those tables.Tue, 07 Jul 2026 00:00:00 GMTsqlitepythondata-engineeringschema-migrationsVikranthttps://vikrantsharma.info/blog/sidenote-llm-writes-git-diffs-from-commentshttps://vikrantsharma.info/blog/sidenote-llm-writes-git-diffs-from-commentsSidenote lets you comment on a rendered blog post, then an LLM writes the actual markdown diff. No forking, no pull requests.Mon, 06 Jul 2026 00:00:00 GMTllmgitdocumentationautomationVikranthttps://vikrantsharma.info/blog/nist-atomic-clock-web-apihttps://vikrantsharma.info/blog/nist-atomic-clock-web-apiThe National Institute of Standards and Technology maintains atomic clocks in Colorado and Maryland, then serves that time over the internet. The API is simpler than you think.Sat, 04 Jul 2026 00:00:00 GMTntptime-syncapinistobservabilityVikranthttps://vikrantsharma.info/blog/mean-field-theory-neural-networkshttps://vikrantsharma.info/blog/mean-field-theory-neural-networksPhysicists found that the standard mean-field approximation for neural networks breaks down when you look at correlations between neurons.Fri, 03 Jul 2026 00:00:00 GMTmachine-learningdeep-learningresearchphysicsVikranthttps://vikrantsharma.info/blog/google-zkp-age-verificationhttps://vikrantsharma.info/blog/google-zkp-age-verificationZKPs let you prove you are over 18 without revealing your birth date. Google just released a library that does it in WebAssembly.Thu, 02 Jul 2026 00:00:00 GMTzero-knowledge-proofscryptographyprivacywebassemblyage-verificationVikranthttps://vikrantsharma.info/blog/ray-tracing-in-clickhouse-sqlhttps://vikrantsharma.info/blog/ray-tracing-in-clickhouse-sqlClickHouse's query language is Turing-complete. Someone proved it by rendering 3D scenes with recursive CTEs and array functions.Wed, 01 Jul 2026 00:00:00 GMTsqldatabasesclickhouseray-tracingdata-engineeringVikranthttps://vikrantsharma.info/blog/micro-agent-beats-frontier-models-with-collaborationhttps://vikrantsharma.info/blog/micro-agent-beats-frontier-models-with-collaborationvLLM's Micro-Agent proves that three coordinated 8B models can outperform a single frontier model on complex reasoning tasks.Tue, 30 Jun 2026 00:00:00 GMTllmreasoningcost-optimizationvllmVikranthttps://vikrantsharma.info/blog/bash-wrapper-llm-apis-no-dependencieshttps://vikrantsharma.info/blog/bash-wrapper-llm-apis-no-dependenciesA GitHub repo called Bash4LLM+ does what Python libraries do in thousands of lines, using only shell builtins and curl.Mon, 29 Jun 2026 00:00:00 GMTllmbashdeveloper-toolsapiVikranthttps://vikrantsharma.info/blog/drones-c4-breach-clearinghttps://vikrantsharma.info/blog/drones-c4-breach-clearingThe US Army tested remote breach clearing with swarms of explosive-packed drones. No soldiers crossed the kill zone.Sun, 28 Jun 2026 00:00:00 GMTdronesmilitary-techroboticsautomationVikranthttps://vikrantsharma.info/blog/lastpass-breach-third-time-this-decadehttps://vikrantsharma.info/blog/lastpass-breach-third-time-this-decadeLastPass notified users of another breach. This is the third major incident since 2022. At what point do we stop calling it a password manager and start calling it a credentials museum.Fri, 26 Jun 2026 00:00:00 GMTcybersecuritypassword-managersdata-breachlastpassVikranthttps://vikrantsharma.info/blog/sqlite-corruption-is-usually-your-faulthttps://vikrantsharma.info/blog/sqlite-corruption-is-usually-your-faultThe SQLite team keeps a public list of every way you can break their database. Turns out most corruption comes from ignoring fsync or writing to the file yourself.Fri, 26 Jun 2026 00:00:00 GMTsqlitedatabasesdata-integrityfilesystemVikranthttps://vikrantsharma.info/blog/haystack-rag-pipelines-not-another-langchainhttps://vikrantsharma.info/blog/haystack-rag-pipelines-not-another-langchainDeepset's Haystack framework caught my eye because it treats retrieval-augmented generation as a data pipeline problem, not a chatbot wrapper problem.Thu, 25 Jun 2026 00:00:00 GMTragmachine-learningpythonvector-searchdata-engineeringVikranthttps://vikrantsharma.info/blog/neural-particle-automata-self-organizing-systemshttps://vikrantsharma.info/blog/neural-particle-automata-self-organizing-systemsResearchers trained particles to form complex shapes without central control. Each particle runs the same neural network, learns local rules, and the swarm organises itself.Wed, 24 Jun 2026 00:00:00 GMTmachine-learningsimulationswarm-intelligenceneural-networksVikranthttps://vikrantsharma.info/blog/tata-electronics-breach-apple-tesla-secretshttps://vikrantsharma.info/blog/tata-electronics-breach-apple-tesla-secretsA ransomware gang says they got into Tata's systems and grabbed confidential files from Apple and Tesla. The supply chain question nobody wants asked.Tue, 23 Jun 2026 00:00:00 GMTcybersecuritysupply-chainransomwareappleteslaVikranthttps://vikrantsharma.info/blog/energy-security-over-climate-goalshttps://vikrantsharma.info/blog/energy-security-over-climate-goalsThe clean energy boom is no longer about emissions targets. It is about not having your grid held hostage by someone else's pipeline.Mon, 22 Jun 2026 00:00:00 GMTenergy-securityrenewable-energycybersecuritygeopoliticsVikranthttps://vikrantsharma.info/blog/amd-drops-memory-encryption-no-warninghttps://vikrantsharma.info/blog/amd-drops-memory-encryption-no-warningA firmware update silently killed Transparent Memory Encryption on consumer chips. The engineers stopped replying when users asked why.Sat, 20 Jun 2026 00:00:00 GMTamdsecurityhardwarefirmwarememory-encryptionVikranthttps://vikrantsharma.info/blog/ten-thousand-fake-github-repos-hosting-trojanshttps://vikrantsharma.info/blog/ten-thousand-fake-github-repos-hosting-trojansSomeone found 10,000 GitHub repositories distributing malware disguised as cracked software and game cheats. The scale is wild.Fri, 19 Jun 2026 00:00:00 GMTsecuritymalwaregithubthreat-intelligenceVikranthttps://vikrantsharma.info/blog/neural-cellular-automata-webgpuhttps://vikrantsharma.info/blog/neural-cellular-automata-webgpuWebGPU makes training tiny neural networks that grow patterns possible in real-time, no server required.Thu, 18 Jun 2026 00:00:00 GMTwebgpumachine-learningcellular-automatabrowserVikranthttps://vikrantsharma.info/blog/databricks-ltap-claims-olap-oltp-unifiedhttps://vikrantsharma.info/blog/databricks-ltap-claims-olap-oltp-unifiedLTAP is Databricks' answer to running transactions and analytics on the same engine. The pitch is clever. The implementation details matter more.Wed, 17 Jun 2026 00:00:00 GMTdatabricksolapoltpdata-engineeringdelta-lakeVikranthttps://vikrantsharma.info/blog/europe-distributed-compute-frontier-modelshttps://vikrantsharma.info/blog/europe-distributed-compute-frontier-modelsA GitHub repo sketches how Europe could pool scattered compute across universities and research labs to train a GPT-4 class model without buying a new datacenter.Tue, 16 Jun 2026 00:00:00 GMTmachine-learningdistributed-systemseuropeinfrastructureVikranthttps://vikrantsharma.info/blog/rio-llm-merge-scandalhttps://vikrantsharma.info/blog/rio-llm-merge-scandalA city government announced a locally trained language model. Turns out it was two existing models stitched together with the weights renamed.Mon, 15 Jun 2026 00:00:00 GMTllmopen-sourcemodel-mergingai-ethicsVikranthttps://vikrantsharma.info/blog/arch-aur-malware-1500-packageshttps://vikrantsharma.info/blog/arch-aur-malware-1500-packagesThe Arch Linux team spent the weekend cleaning up more than 1,500 malware-laden packages from the AUR. The scale is what surprised me.Sun, 14 Jun 2026 00:00:00 GMTlinuxsecuritysupply-chainpackage-managementVikranthttps://vikrantsharma.info/blog/arch-aur-malware-compromise-2026https://vikrantsharma.info/blog/arch-aur-malware-compromise-2026The Arch User Repository just had 400 packages backdoored. The attack vector was not clever. It was obvious, predictable, and it worked anyway.Sat, 13 Jun 2026 00:00:00 GMTlinuxsecuritysupply-chainarch-linuxVikranthttps://vikrantsharma.info/blog/ivanti-sentry-perfect-cvss-scorehttps://vikrantsharma.info/blog/ivanti-sentry-perfect-cvss-scoreIvanti Sentry got a pre-auth RCE with the maximum theoretical severity score. Public exploit code is already live.Fri, 12 Jun 2026 00:00:00 GMTcybersecurityvulnerability-managementivanticveexploitVikranthttps://vikrantsharma.info/blog/transformer-attention-executive-control-deficithttps://vikrantsharma.info/blog/transformer-attention-executive-control-deficitResearchers found that transformer attention mechanisms lack the executive control functions that let human brains manage working memory. The models can retrieve information, but they cannot suppress irrelevant context.Thu, 11 Jun 2026 00:00:00 GMTtransformersmachine-learningcognitive-scienceattention-mechanismsVikranthttps://vikrantsharma.info/blog/servicenow-breach-configuration-not-codehttps://vikrantsharma.info/blog/servicenow-breach-configuration-not-codeThe ServiceNow incident exposed customer data through misconfigured access controls, not a novel exploit. This is the enterprise security story that never makes headlines.Wed, 10 Jun 2026 00:00:00 GMTcybersecuritysaasaccess-controlservicenowconfigurationVikranthttps://vikrantsharma.info/blog/config-files-that-execute-codehttps://vikrantsharma.info/blog/config-files-that-execute-codeYAML, TOML, and JSON parsers can run arbitrary code during deserialization. Most dependency scanners miss them entirely.Tue, 09 Jun 2026 00:00:00 GMTsupply-chainyamlsecuritydependenciesVikranthttps://vikrantsharma.info/blog/oxford-career-platform-breach-third-party-riskhttps://vikrantsharma.info/blog/oxford-career-platform-breach-third-party-riskThird-party integrations are the soft underbelly of university IT. Oxford just learned that again.Sun, 07 Jun 2026 00:00:00 GMTthird-party-riskdata-breachsaas-securityuniversity-itaccess-managementVikranthttps://vikrantsharma.info/blog/meta-llama-4-delay-reasoning-modelshttps://vikrantsharma.info/blog/meta-llama-4-delay-reasoning-modelsMeta keeps pushing back the Llama 4 release because the reasoning capability is not matching internal benchmarks. This is the first time open-weight AI has hit a public delay.Sat, 06 Jun 2026 00:00:00 GMTllamametareasoning-modelsopen-weightsai-infrastructureVikranthttps://vikrantsharma.info/blog/sqlite-uuid-primary-keys-performance-hithttps://vikrantsharma.info/blog/sqlite-uuid-primary-keys-performance-hitRandom UUIDs as primary keys in SQLite cause page splits that triple insertion time. Integer keys stay fast because SQLite is built for sequential writes.Sat, 06 Jun 2026 00:00:00 GMTsqlitedatabase-performanceuuidindexingVikranthttps://vikrantsharma.info/blog/anthropic-released-a-vuln-discovery-harnesshttps://vikrantsharma.info/blog/anthropic-released-a-vuln-discovery-harnessAnthropic open-sourced a framework for testing how well AI models find security bugs. It includes 32 real CVEs and a scoring system. Time to feed it some of my old projects.Fri, 05 Jun 2026 00:00:00 GMTai-securityvulnerability-researchanthropicbenchmarkingcode-analysisVikranthttps://vikrantsharma.info/blog/preview-urls-without-kuberneteshttps://vikrantsharma.info/blog/preview-urls-without-kubernetesA self-hosted tool that spins up Docker containers with public URLs. No orchestrator, no cloud bill.Thu, 04 Jun 2026 00:00:00 GMTdockerdevopsself-hostedpreview-environmentsVikranthttps://vikrantsharma.info/blog/data2prompt-stuffs-projects-into-llm-contexthttps://vikrantsharma.info/blog/data2prompt-stuffs-projects-into-llm-contextA CLI tool that flattens your data science repo into one massive prompt. Smart filtering meets the 200K token era.Wed, 03 Jun 2026 00:00:00 GMTmachine-learningllmtoolingclicode-reviewVikranthttps://vikrantsharma.info/blog/python-becoming-pinyinhttps://vikrantsharma.info/blog/python-becoming-pinyinReuven Lerner argues Python's syntax sugar is making it harder to explain what code actually does. He compares it to Pinyin romanisation: useful for getting started, obscuring the underlying structure.Tue, 02 Jun 2026 00:00:00 GMTpythonprogramming-languagesteachingVikranthttps://vikrantsharma.info/blog/chatgpt-google-sheets-leakhttps://vikrantsharma.info/blog/chatgpt-google-sheets-leakA popular extension with 3 million users can be tricked into sending your spreadsheet contents to an attacker-controlled server. The fix is not obvious.Sun, 31 May 2026 00:00:00 GMTprompt-injectionllm-securitygoogle-sheetsdata-exfiltrationVikranthttps://vikrantsharma.info/blog/optane-memory-trillion-parameter-llmhttps://vikrantsharma.info/blog/optane-memory-trillion-parameter-llmAn enthusiast loaded a 1T-parameter model into 768GB of Intel Optane DIMMs and got 4 tokens per second on a single GPU. Slow, but it worked.Sun, 31 May 2026 00:00:00 GMTllmhardwareoptaneinferencememoryVikranthttps://vikrantsharma.info/blog/end-to-end-encrypted-home-cameras-without-the-cloud-taxhttps://vikrantsharma.info/blog/end-to-end-encrypted-home-cameras-without-the-cloud-taxSomeone finally built a home security camera system that encrypts on-device and costs nothing per month.Sat, 30 May 2026 00:00:00 GMTsecurityprivacyself-hostedcomputer-visionVikranthttps://vikrantsharma.info/blog/llm-code-smells-hedging-and-fillerhttps://vikrantsharma.info/blog/llm-code-smells-hedging-and-fillerA developer catalogues the tell-tale signs of AI-generated code. The patterns are obvious once you see them.Fri, 29 May 2026 00:00:00 GMTllmcode-reviewdetectionsoftware-engineeringVikranthttps://vikrantsharma.info/blog/composer-package-signing-finally-arriveshttps://vikrantsharma.info/blog/composer-package-signing-finally-arrivesPHP's dependency manager is rolling out cryptographic signatures. The timing matters more than the tech.Thu, 28 May 2026 00:00:00 GMTsupply-chainphpsecuritypackage-managementVikranthttps://vikrantsharma.info/blog/cloudflare-flagship-feature-flags-at-the-edgehttps://vikrantsharma.info/blog/cloudflare-flagship-feature-flags-at-the-edgeFeature flags that run on Cloudflare's edge network, not your backend. This might actually change how you roll out features.Wed, 27 May 2026 00:00:00 GMTcloudflarefeature-flagsedge-computingdevopsVikranthttps://vikrantsharma.info/blog/norway-huawei-flash-llm-traininghttps://vikrantsharma.info/blog/norway-huawei-flash-llm-trainingThe Norwegian University of Science and Technology is running LLM workloads on Chinese storage hardware. The performance numbers are interesting.Tue, 26 May 2026 00:00:00 GMTmachine-learningstoragegpuinfrastructureVikranthttps://vikrantsharma.info/blog/linux-packet-fragmentation-exploitshttps://vikrantsharma.info/blog/linux-packet-fragmentation-exploitsDirty Frag, Copy Fail, Fragnesia. Three distinct kernel exploits in eighteen months, all exploiting how Linux reassembles fragmented network packets.Sun, 24 May 2026 00:00:00 GMTlinuxkernel-securitynetworkingexploitpacket-fragmentationVikranthttps://vikrantsharma.info/blog/llm-agents-forget-constraints-when-generating-backend-codehttps://vikrantsharma.info/blog/llm-agents-forget-constraints-when-generating-backend-codeNew research shows agents generating backend code slowly drop requirements like authentication checks. The longer the generation, the worse the decay.Sun, 24 May 2026 00:00:00 GMTllmcode-generationbackendconstraintsagentsVikranthttps://vikrantsharma.info/blog/models-dev-database-ai-model-specshttps://vikrantsharma.info/blog/models-dev-database-ai-model-specsModels.dev is an open-source database that tracks pricing, context windows, and rate limits across every major LLM provider. No more tab-sprawl to compare GPT-4 versus Claude costs.Fri, 22 May 2026 00:00:00 GMTllmpricingopen-sourceaitoolingVikranthttps://vikrantsharma.info/blog/content-moderation-becomes-federal-lawhttps://vikrantsharma.info/blog/content-moderation-becomes-federal-lawA bill requiring platforms to moderate content encouraging violence against Jewish communities turns moderation from a platform choice into a legal obligation.Thu, 21 May 2026 00:00:00 GMTcontent-moderationpolicyplatform-governancemachine-learningVikranthttps://vikrantsharma.info/blog/shira-phishing-training-gamifies-detectionhttps://vikrantsharma.info/blog/shira-phishing-training-gamifies-detectionA new platform gamifies phishing detection with streak tracking and leaderboards. Finally, security awareness training that does not feel like compliance homework.Thu, 21 May 2026 00:00:00 GMTcybersecurityphishingsecurity-traininggamificationVikranthttps://vikrantsharma.info/blog/ai-bug-hunters-flooding-linux-securityhttps://vikrantsharma.info/blog/ai-bug-hunters-flooding-linux-securityLinus Torvalds says automated vulnerability scanners have turned the kernel security mailing list into noise. The tools work, the signal-to-noise ratio does not.Tue, 19 May 2026 00:00:00 GMTlinuxsecuritymachine-learningopen-sourceVikranthttps://vikrantsharma.info/blog/formal-verification-found-sqlite-bugshttps://vikrantsharma.info/blog/formal-verification-found-sqlite-bugsTurso used a TLA+ variant called Quint to model their libSQL fork and found over ten real bugs in SQLite's transaction logic that billions of existing tests never saw.Tue, 19 May 2026 00:00:00 GMTformal-verificationsqlitetestingdistributed-systemsVikranthttps://vikrantsharma.info/blog/uk-vpn-regulation-mozilla-pushbackhttps://vikrantsharma.info/blog/uk-vpn-regulation-mozilla-pushbackThe Online Safety Bill could give UK authorities the power to demand VPN providers weaken encryption or log traffic. Mozilla's response is blunt.Mon, 18 May 2026 00:00:00 GMTvpnprivacyregulationencryptionuk-policyVikranthttps://vikrantsharma.info/blog/deepseek-v4-flash-steering-vectorshttps://vikrantsharma.info/blog/deepseek-v4-flash-steering-vectorsSteering vectors let you nudge a model's behaviour without retraining. They fell out of favour when newer models stopped responding to them. DeepSeek-V4-Flash brought them back.Sun, 17 May 2026 00:00:00 GMTmachine-learningllmdeepseekinterpretabilityVikranthttps://vikrantsharma.info/blog/rust-category-theory-ml-frameworkhttps://vikrantsharma.info/blog/rust-category-theory-ml-frameworkA programmer rewrote PyTorch's transformer architecture using Rust and abstract algebra. The result is dense but fast.Fri, 15 May 2026 00:00:00 GMTrustmachine-learningcategory-theorytransformerspytorchVikranthttps://vikrantsharma.info/blog/m5-kernel-exploit-publichttps://vikrantsharma.info/blog/m5-kernel-exploit-publicThe first public kernel memory corruption exploit for Apple's M5 chip dropped today. No zero-day trading, just research in the open.Thu, 14 May 2026 00:00:00 GMTapple-siliconkernel-exploitsmacossecurity-researchm5Vikranthttps://vikrantsharma.info/blog/torrix-observability-without-the-database-taxhttps://vikrantsharma.info/blog/torrix-observability-without-the-database-taxA new LLM observability tool runs without PostgreSQL or Redis. That is not a feature list, that is an architecture decision.Thu, 14 May 2026 00:00:00 GMTllmobservabilitysqliteself-hostedVikranthttps://vikrantsharma.info/blog/anthropic-claude-aws-marketplacehttps://vikrantsharma.info/blog/anthropic-claude-aws-marketplaceClaude's API is now inside AWS Marketplace. That means billing through your AWS account, usage tied to AWS credits, and regional deployment closer to your data.Wed, 13 May 2026 00:00:00 GMTawsanthropicclaudemlopscloudVikranthttps://vikrantsharma.info/blog/ai-finds-cve-in-its-own-training-sethttps://vikrantsharma.info/blog/ai-finds-cve-in-its-own-training-setMythos discovered a vulnerability that was already documented in the data it was trained on. The industry is calling this autonomous discovery.Mon, 11 May 2026 00:00:00 GMTai-securityllmvulnerability-researchcvemachine-learningVikranthttps://vikrantsharma.info/blog/curl-vulnerability-found-by-llm-agenthttps://vikrantsharma.info/blog/curl-vulnerability-found-by-llm-agentMythos, an autonomous security agent, caught a buffer overflow in curl that human auditors missed. The tooling works.Mon, 11 May 2026 00:00:00 GMTsecurityllmopen-sourcecurlvulnerabilityVikranthttps://vikrantsharma.info/blog/gemini-api-file-search-goes-multimodalhttps://vikrantsharma.info/blog/gemini-api-file-search-goes-multimodalGoogle's Gemini API File Search now handles images inside documents. RAG just got less annoying for technical docs.Mon, 11 May 2026 00:00:00 GMTgeminiragmultimodalgoogleaiVikranthttps://vikrantsharma.info/blog/llm-agents-should-not-talk-to-each-other-in-englishhttps://vikrantsharma.info/blog/llm-agents-should-not-talk-to-each-other-in-englishMulti-agent systems waste tokens on natural language between agents. A structured clipboard beats conversational interfaces.Mon, 11 May 2026 00:00:00 GMTllm-agentsmulti-agent-systemstoken-efficiencysystem-designVikrant