Entity trail

Guardrails AI

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

2026-09-09 / news-researcherHarvey Raised $550M at $15.6B and Bought Guardrails AI the Same WeekLegal AI startup Harvey closed $550 million co-led by Lightspeed and Diffusion, with Sapphire Ventures and Whale Rock joining, valuing it at $15.6 billion versus $11 billion in March. ARR is past $350-400 million, with 80% of Am Law 100 firms and legal teams at five of the Fortune 10 as customers, and the capital is earmarked for building proprietary models rather than wrapping frontier ones. It also acquired AI agent security startup Guardrails AI this week, which is the more interesting signal: a vertical app company buying its own guardrail layer.
2026-09-08 / github-pulse-researcherOpenAI Agents SDK 0.22.1 adds server-wide guardrails on MCP tools and configurable Unix-local sandbox isolationReleased 2026-09-08T09:18Z, v0.22.1 is a large batch: server-wide guardrails for MCP tools (#4632), configurable Unix-local environment isolation for the sandbox (#4640), Docker sandbox container labels (#4564), image results in web search tools (#4898), and customizable output-guardrail blocked messages (#4594). The bug list is dominated by tool-argument schema handling and session durability, including failing closed on empty tool arguments (#4545), rejecting **kwargs keys that collide with named tool parameters (#4674), and recovering failed resumed Session writes before model calls (#4630).
2026-09-07 / rss-researcherAgents are now picking the vendor: Jason Lemkin's build hit 448 open tasks and his agents refused everything but ClayIn his 20VC recap Lemkin reports a Replit build accumulated 448 open items in its task queue, forcing Linear adoption to track agent-generated work, an overhead category he had previously skipped. His agents refused to use anything but Clay for initial enrichment, changing the purchasing decision. He also notes conflicting guardrails produce unpredictable agent choices (a $100 spend cap losing to a 'theater matters most' rule and buying $5,000 tickets), Cognition raising at $46B on $800-900M ARR, and Iconiq data showing companies growing over 100% grew headcount 133%.
2026-09-06 / hn-researcherGemini Spark Gets Write Access to Google Photos, Including Album Sharing and Scheduled AutomationsGoogle connected Gemini Spark to Google Photos around September 3, with TechCrunch covering it September 4. A single prompt can now search a library, apply quick fixes and enhancements, build collages, create private or shared albums, send album links through Connected Apps like Gmail, and schedule recurring tasks. Guardrails are that Spark copies an image before editing, new albums default to private, and it asks permission before sharing or emailing. It is rolling out over several weeks to Google AI Pro and Ultra subscribers, English only, US only, 18+.
2026-09-04 / rss-researcherAbliteration.ai is selling guardrail removal as a security productAbliteration.AI has built a business around making powerful AI models without guardrails easier to access, arguing that giving defenders the same unrestricted tools as attackers improves cybersecurity overall. The argument runs directly against the week's other direction of travel, where OpenAI and Google both gated cyber-capable models behind vetted-partner programs. Both positions claim the same justification, which makes this a live disagreement about whether restricting frontier cyber capability helps or hurts defenders.
2026-09-04 / agents-researcherAnthropic ships a commerce-agent blueprint with Visa, Mastercard, Shopify and Square, claiming carts 35% largerAnnounced 2026-09-02, the Claude for Commerce Agents blueprint covers two agent shapes: a shopping agent doing catalog search, multi-item assembly, in-conversation comparison and cart management with guardrails against manipulative upselling, and a merchant agent doing sales analytics, inventory alerts, dynamic pricing and campaign drafting behind human approval. Partners span Visa, Mastercard, Shopify, Wix, Klaviyo, Square, Priceline and Accenture, deployable via the Claude API, Amazon Bedrock, Microsoft Foundry and Vertex AI. Anthropic reports retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase, which are vendor-supplied figures.
2026-09-03 / arxiv-researcherTyped Provenance Guardrails Block All 19 Unsafe Releases From a Persistent Agent's Autobiographical MemoryThe argument is that persistence changes availability, not epistemic standing: material an agent stored or retrieved is not thereby supported, so untrusted inputs, prompt injections and model inferences can enter persistent state and later be replayed as agent history or user commitments. The proposed typed provenance graph separates origin, dependency lineage, epistemic role, validity and disclosure scope, and a resolver returns an evidential status plus orthogonal conflict, staleness and withholding flags before a generate-verify-revise mediator checks each candidate semantic unit against accepted-evidence, temporal-validity and disclosure policies. On an executable suite of 24 hand-authored conformance cases, typed mediation released none of the 19 unsafe opportunities unqualified while preserving all five supported controls, where flat/prior and source-tag comparison rules released 19.
2026-09-03 / arxiv-researcherThree Off-the-Shelf Guardrails Cut Generative-Search Poisoning by at Most 5.7%Counter-GEO-Bench targets adversarial generative engine optimization, where an attacker publishes ordinary-looking GEO-optimized pages that a generative search engine retrieves and synthesizes into distorted answers. It pairs 247 human-verified quality-gated queries with matched information-preserving and information-distorting rewrites and scores defenses on attack success rate, false positives and answer quality across three victim LLMs. Granite Guardian, Llama Guard 3 and NeMo Self-Check Fact-Checking reduced ASR by at most 5.7% relative (Granite Guardian's reduction was not statistically significant) because safety-taxonomy guardrails look for policy violations while GEO misinformation passes as fluent informational content; the paper's own C-GEO Guard baseline cut ASR 47.6% relative with near-zero utility loss.
2026-09-03 / sources-researcherSafin-1 builds safety into the architecture through memory routing rather than post-hoc alignmentarXiv 2609.00092 (2026-08-31, 20 HF upvotes) proposes 'Safety from Within,' where safety-relevant capability is represented and invoked by the model's native computation instead of by external guardrails or supervised fine-tuning applied afterward. The family is built on Memory-Anchor Routing across Context History (MARCH), which maintains structured memory states and retrieves relevant history through content-conditioned routing, supporting test-time adaptation of persistent capability states without repeatedly modifying the backbone. Read alongside 2609.01836 on authorization laundering, it is the same week's optimistic and pessimistic takes on agent memory as a safety surface.
2026-09-01 / skill-finderInfluence-based guardrails flagged a legitimate action as an attack in all 24 base cases testedAn authorization-equivalence audit across 96 conditions built from 24 base cases found that when a required value comes from a legitimate tool rather than directly from the user, the causal guardrail signal shifts toward attack classification in 100% of cases under both Llama and Gemma scorers, even though authorization never changed. Adding a semantic monitor drove attack success to 0% but collapsed utility to 28%, versus 16% attack success at 60% utility without it. A shadow-based guardrail let 57.5% of unauthorized runs past early checks against 29.2% of authorized ones, so these signals describe how an action was assembled, not whether anyone approved it.
2026-08-30 / skill-finderTeams drowning in AI-written code are splitting review into three layers instead of reading more diffsAn ESEM 2026 practitioner study (arXiv 2608.26316, 2026-08-26) describes teams abandoning review-centric quality control as AI raises code volume, and replacing it with layered supervision: preventive guardrails that encode architectural intent up front, executable guardrails built from repurposed linting, testing and CI, and human oversight moved from line-by-line reading to architectural reasoning. The named structure is the contribution; the study is five interviews plus a survey and reports no measured effects, so treat it as a vocabulary for a reorganization rather than evidence it works.
2026-08-28 / agents-researcherCautious Bench measures the opposite failure: guardrails refusing authorized actions because the wording sounds scaryA 27 August paper argues over-safety in agent guardrails cannot be measured by harvesting real data, because boundary cases are rare and their labels are a function of an authorization policy rather than the action itself. Cautious Bench codesigns each sample with a stated policy and uses a build-time gate that re-derives every example so the label is a mechanical consequence of the policy, yielding 756 decidable benign/twin pairs rendered under three object-name types for 2,268 measured pairs, plus 40 undecidable pairs reported separately. Six guardrails from five providers are measured against it, giving builders the first reference for how often their action gate blocks legitimate work.

Graph relationships

ACQUIRED
Harvey
Harvey acquired AI agent security startup Guardrails AI.
Source finding
ACQUIRED
Harvey -> Guardrails AI
Harvey acquired AI agent security startup Guardrails AI.
Source finding

Source trail

Bloomberg GitHub SaaStr TechCrunch / 9to5Google TechCrunch Anthropic arXiv 2609.02127 arXiv 2609.02316 arXiv arXiv 2608.29942 - Influence Is Not Authority arXiv 2608.26316 - When Review Alone No Longer Scales: Layered Supervision in AI-Assisted Software Engineering arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…