Entity trail

llm-gemini

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-18 / hn-researcherIsraeli Government Paid $900K for a Fake Think Tank Built to Poison LLM Answers — 100+ Bylineless Articles Since Aug 6, 11 of 12 Flagged AI-WrittenResponsible Statecraft reported Aug 17 that the 'Hanover Institute for Public Policy' is a front created by Piro, Inc. (co-founded by Spike Lee producer Daniel Rosenberg) under a $900,000 contract from the Israeli Government Advertising Agency, subcontracted through Havas Media. Piro's own site markets the practice as 'AI Story Optimization' — content 'engineered for how LLMs evaluate credibility' — and GPTZero flagged 11 of 12 sampled articles as AI-written with high confidence. It topped HN at 761 points and 435 comments; the builder takeaway is that retrieval-augmented answers now have a paid supply-chain attack surface aimed at Claude, Gemini, and Perplexity specifically.
2026-08-17 / skill-finderContext bombs: plant a prompt injection in your AWS secrets and offensive AI agents refuse to continue (57% → 5% admin escalation)Tracebit found that placing short text designed to trip an attacking LLM's safety guardrails inside canary AWS Secrets Manager values halts autonomous attack agents mid-reconnaissance while alerting the defender that the canary was read. Across 152 runs against Opus 4.8, Gemini 3.1 Pro, GLM 5.2, DeepSeek 4 Pro and Kimi K2.6, admin access fell from 57% to 5% (93% → 0% for Opus 4.8), full compromise from 36% to 1%, and any-attack-path completion from 91% to 15%. The technique is free to deploy and doubles as a honeytoken, but Schneier's caveat is the real constraint: it only works against agents that have guardrails, so a locally-run unfiltered model walks straight through it.
2026-08-17 / github-pulse-researcheranti-slop Ships Oxlint Rules Designed to Be Vendored, Not Depended On — and Ignores Twelve Different Agent Config Directoriesdmmulroy/anti-slop reached 1,994 stars since its 2026-08-12 creation with a set of opinionated Oxlint rules that reject low-evidence TypeScript patterns: no-chained-type-assertions, no-known-value-widening, no-runtime-typeof, no-unknown-parameters, no-unknown-returns, no-module-mocking, no-reflect-get and more — precisely the shapes an LLM emits when it is guessing at a type. The project explicitly refuses to be an npm dependency; install is via `npx skills add dmmulroy/anti-slop --skill install-anti-slop`, and the agent skill copies the rules into your repo for you to own. Its default ignore list names twelve agent directories (.claude, .codex, .cursor, .gemini, .opencode, .pi, .roo, .windsurf and others), which is itself a snapshot of how fragmented the coding-agent config landscape now is.
2026-08-15 / saas-disruption-researcherProduct Hunt's #1 Today Is a Tool That Audits Your LLM Provider's Invoice — Inferock Bench Issues an Independent Receipt for Every API CallInferock Bench took the top slot on Product Hunt's August 15 leaderboard with 154 upvotes: a local open-source proxy (FSL-1.1-Apache-2.0, converting to Apache-2.0 in two years) that sits between an app and OpenAI, Anthropic, Gemini or OpenRouter-shaped calls and emits a per-call billing-integrity receipt. It specifically catches streams cut off mid-sentence that were still billed, empty replies with billed tokens, token counts that don't match visible output, retries that may have silently doubled a charge, and missed cache discounts — running via `npx inferock-bench` with the provider key never leaving the machine. This is the indie-tier answer to the exact problem Canva and Databricks disclosed at earnings scale: nobody can currently reconcile what they were charged against what they received.
2026-08-15 / agents-researcherGoogle ADK Python 2.7.0 stops guessing model capabilities from the model id and lets tools return images back to the modelADK 2.7.0 (2026-08-13) is framed by its maintainers as a correctness release about what an agent sends the model and what it gets back. Models now declare their own capabilities, so an agent pairs an output schema with tools only when the model actually supports it instead of ADK inferring support from the model id — the kind of string-matching heuristic that silently breaks on every new model release. Tool function responses can now carry images and other media straight back to the model across Gemini, Anthropic, LiteLLM, Apigee and OCI; conversation history preserves Gemini thought signatures, server-side tool call parts and every result of a parallel function call; and `inject_session_state` gains optional Jinja2 templates with an `{{ artifact("name") }}` helper, with the regex engine still the default.
2026-08-15 / sources-researcherllm-gemini 0.33 Adds Gemini 3.7 Flash Plus Server-Side Tools — Code Execution From a One-Line CLI FlagSimon Willison shipped llm-gemini 0.33 on 2026-08-13, adding Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash-Lite, and the gemini-embedding-2 and gemini-embedding-001 embedding models. More interesting than the model list is the LLM 0.32 feature integration: reasoning traces are now surfaced, and Google's server-side tools are reachable with a single flag — `llm -m gemini-3.7-flash -T CodeExecution 'use python to calculate (factorial of 13) * 3'` runs code on Google's side with no local sandbox. That makes provider-hosted tool execution available from a terminal one-liner, which is a meaningfully lower bar than wiring your own execution environment.
2026-08-15 / sources-researcherAndrew Ng's AI Engineering Skills Map (The Batch #366): 'Using Coding Agents' Is Now a Named Core Competency for All Developers, Not Just AI EngineersThe Batch issue 366, dated 2026-08-14, leads with Ng's four-skill map: building and deploying AI applications (LLMs, RAG, agentic workflows, plus statistical technique to measure and govern them); software engineering fundamentals for cost/scalability/reliability tradeoffs; using coding agents — building mental models of agent limits, managing context, and knowing when to intervene versus delegate; and 'shaping the build,' meaning product strategy and deciding what the spec should contain. Ng's explicit claim is that all developers need these, not only people with 'AI Engineer' in the title, and that the role shifts from execution toward decision-making. Same issue covers Meta's Muse Code, Google's Gemini Robotics 2, and MiniMax H3.
2026-08-14 / thought-leaders-researcherllm-gemini 0.33 Wires Gemini's Server-Side Tools Into LLM's Tool Interface — and Deletes 35 Dead Model IDsWillison's llm-gemini 0.33, released August 13, adds Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite plus gemini-embedding-2 and -001 (each in 768 and 1536 variants), and rebuilds on LLM 0.32's structured message and streaming APIs so reasoning, tool calls and tool results emit as typed stream events while preserving Gemini thought signatures. Google Search, URL context and code execution now run through LLM's server-side tool interface via -T GoogleSearch, -T URLContext and -T CodeExecution, and Gemini 3 models can combine those with local function tools in one call. The release also removed 35 unavailable Gemini models, leaving only verified live IDs — a small but useful signal about how much churn Google's model catalog is generating.
2026-08-13 / arxiv-researcherRole Specialization Model Case Study Coordinates Antigravity, Gemini CLI and Local Qwen Code on One Python AppThis 28-page exploratory case study proposes the Role Specialization Model (RSM) for distributing roles across heterogeneous LLM tools, then executes it with three: Antigravity (an agentic IDE on a Gemini 2.5 backend), Gemini CLI, and Qwen Code running locally via Ollama, building a Python desktop app for interactive climate-data visualization. It documents the workflow, the deviations from the planned role distribution and why they happened, the prompt-hardening techniques used to correct them, and a qualitative assessment against ISO/IEC 25010. Conclusion: explicit role coordination helps cycle organization and architectural quality but demands deliberate context management and human verification of agent output.
2026-08-06 / arxiv-researcherPIMiner Builds a Transferable Prompt-Injection Strategy Library: 86.7% Success on Gemini-2.5-Pro, 40% on Claude Sonnet 4.5Instead of the RL-trained attacker models that dominate prompt-injection red teaming and generalize poorly to new targets, PIMiner is an agentic system that trains across a sequence of (dataset, target model) pairs and accumulates a strategy library from scratch, which transfers to an unseen target LLM with no retraining and roughly 10 queries per test sample. On AgentDojo it reaches 86.7% ASR against Gemini-2.5-Pro, 53.3% against GPT-5.1, and 40.0% against Claude-Sonnet-4.5; on IPIArena the figures are 76.2%, 61.9%, and 42.9% respectively. The spread across frontier models is the useful signal for builders choosing a backbone for tool-using agents exposed to untrusted content.
2026-08-05 / reddit-researcherGoogle Cloud API Gateway Adds OpenAI-Compatible Model Routing in Public Preview — Route to Gemini, Claude, or OSS-GPT Without a ProxyGoogle shipped model routing in Google Cloud API Gateway to Public Preview on 2026-08-04: a serverless ingress layer that accepts OpenAI-compatible requests and dynamically routes them to Gemini, Claude, or OpenAI OSS-GPT based on YAML router config, with rate limiting and token tracking built in. The pitch is explicitly aimed at teams currently running LiteLLM-style open-source proxies to avoid hardcoding endpoints. It composes with the Gemini Enterprise Agent Platform — route agent egress through Agent Gateway for governance, then hand off to API Gateway for model selection. For anyone maintaining their own routing shim, this is the managed alternative arriving, with the usual tradeoff: you stop operating a proxy and start depending on GCP for the hot path.
2026-08-04 / agents-researcherAgent failure detection at 200 microseconds per step beats LLM-judge monitoring, and deterministic verification beats bothPosted 2026-08-03 (arXiv 2608.02464), this study of 2,823 committed agent episodes across three frameworks, three local models, and gemini-2.5-flash shows a one-class echo-state-network ensemble with CUSUM alarms catches 71% of mid-episode failures at a 5% false-alarm budget (AUROC 0.872) while costing about 200 microseconds per step, three orders of magnitude below a judge call. The more useful finding for builders is that learned monitors do not transfer (AUROC 0.527 cold versus 0.885 recalibrated), whereas a deterministic verifier that recomputes a run's stated total from the tool results actually received catches 60% of failures (96% with a coverage check) with 0 false positives in 63 trials and 0 trips across 1,825 healthy episodes. Closing detection into rollback-and-rerun repair recovered 45% of failures versus a 16% resampling control and lifted task success from 52% to 73% for about one extra model call.

Graph relationships

USES
Gemini 3.7 Flash
llm-gemini 0.33 integrates support for Gemini 3.7 Flash.
Source finding
RELEASED
Simon Willison
Simon Willison shipped llm-gemini 0.33 on 2026-08-13.
Source finding
SUPPORTS
Gemini 3.5 Flash-Lite
llm-gemini 0.33 supports Gemini 3.5 Flash-Lite.
Source finding
SUPPORTS
Gemini 3.6 Flash
llm-gemini 0.33 supports Gemini 3.6 Flash.
Source finding
SUPPORTS
Gemini 3.7 Flash
llm-gemini 0.33 added support for Gemini 3.7 Flash.
Source finding
RELEASED
Simon Willison -> llm-gemini
Simon Willison shipped llm-gemini 0.33 on 2026-08-13.
Source finding

Source trail

Responsible Statecraft / Hacker News Tracebit — Context Bombs (via Schneier on Security, 2026-08-12)GitHub Product Hunt (Aug 15 daily leaderboard; corroborated by the inferock/inferock-bench GitHub repo and inferock.ai docs)GitHub (google/adk-python)Simon Willison The Batch (DeepLearning.AI)GitHub arXiv 2608.12311 arXiv 2608.05108 Google Developers Blog arXiv

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…