Daily AI Digest

Perplexity launches 'Computer for Counsel' to crack the legal market

Perplexity states its competitive edge centres on accuracy: 'most LLMs tolerate a certain amount of hallucinations...at Perplexity, we only reward answers that are factually accurate.'

Harvey develops open-source legal foundation models for firm-owned intelligence

Co-founder Gabe Pereyra: Harvey aims to enable 'law firms to build their own specialized models and own their own intelligence,' developing 'the first model in our legal foundation model series' to deliver intelligent capabilities affordably while maintaining security.

Ironclad and Legora build bidirectional AI-to-AI integration inside each other's platforms

Ironclad CEO Dan Springer: the integration enables 'AI systems that work together' for faster regulatory analysis, litigation assessment, and contract impact determination — all within trusted enterprise systems.

Kirkland's $500M Palantir deal is mostly a bet on data architecture, not AI software

Ken Crutchfield: 'The goal is not simply to generate better drafts. It is to transform decades of experience into a strategic asset.' He compares it to Shearman & Sterling's years-long project tagging a billion documents — unglamorous work that proved essential for effective AI deployment.

DiffusionGemma generates 256 tokens at once — 4x throughput on consumer GPU hardware

DiffusionGemma 'generates entire blocks of text simultaneously' via text diffusion rather than sequential token prediction, achieving 1,000+ tokens/second on NVIDIA H100 and 700+ on RTX 5090.

GLM-5.2 tops open-weights benchmarks with 753B MoE and 1M token context

Simon Willison: GLM-5.2 is 'probably the most powerful text-only open weights LLM,' leading the Artificial Analysis Intelligence Index v4.1 at score 51 and ranking 2nd on the Code Arena WebDev leaderboard — behind only Claude Fable 5.

Thomson Reuters ships CoCounsel + DeepJudge integration for firm knowledge search

DeepJudge co-founder Paulina Grnarova: the integration 'surfaces relevant institutional knowledge before AI model processing, reducing token consumption' and works across 'complex legal repositories with multiple versions, duplicates, and strict ethical walls' without requiring firms to migrate their documents.

BlackBoiler Veris pairs deterministic redlining with generative AI inside Microsoft Word

CEO Daniel Broderick: Veris aims to 'only edit what needs to be edited and not more than what needs to be edited.' A 'judge' component validates AI-suggested changes against the firm's historical contract edits before inserting them as tracked changes.

Relativity acquires Gavel to connect case-strategy AI with Word drafting

Relativity's Chief Product Officer: 'We would be taking the system of action that lawyers already rely on and extending it' into Microsoft Word — allowing work product from aiR for Case Strategy and aiR Assist to be opened, drafted, and redlined in Word with changes syncing back to RelativityOne.

Berkeley Law AI ban draws student call for disclosure model over prohibition

LL.M. student Lakshita Bhargava: 'Instead of prohibition, law schools should adopt a permitted-but-disclosed model where students openly report AI usage.' She notes the ban 'affects even the Master of Laws with AI Governance focus' — making the restrictions internally inconsistent with the school's own academic mission.

EU AI Act GPAI enforcement powers activate in seven weeks — fines now a real risk

Starting August 2, 2026, the Commission can request documentation, conduct independent evaluations, demand compliance measures and risk mitigation, restrict market availability, require recalls or forced withdrawal, and impose fines up to 3% of annual worldwide turnover or €15 million, whichever is higher.

US export control order suspends all access to Fable 5 and Mythos 5

Anthropic received the directive on June 12 at 5:21 PM ET. The government cited 'a method of bypassing, or jailbreaking Fable 5' — specifically, asking the model 'to read a specific codebase and fix any software flaws.' Anthropic notes this capability 'is widely available from other models' and used daily by security professionals.

Anthropic Public Record: only 15% of Americans trust AI companies

Only 15% of Americans trust AI companies to make development decisions — the lowest institutional trust rating tested, far below independent experts at 43%. Over 70% believe government should regulate AI, with support spanning 71% of Democrats, 68% of Republicans, and 69% of Independents.

TCS-Anthropic deal deploys Claude to 50,000 employees across 56 countries

TCS will deploy Claude across 50,000 of its own employees in 56 countries and to thousands of enterprise clients globally, with a focus on financial services, healthcare, public sector, life sciences, aviation, telecom, and medical technology.

'Open source AI must win' manifesto draws 1,000+ points on Hacker News

Ahmad Osman: 'When intelligence becomes something people can only rent from a few closed institutions, we lose more than software freedom.' He frames AI as civilizational infrastructure that must remain 'usable, understandable, reproducible, locally deployable, economically viable, and community-governed.'

Five legal tech companies name new CEOs in a single week

Bob Ambrogi: 'This unusual concentration of leadership changes' — four new CEOs at 8am (parent of LawPay/MyCase), Casepoint, Steno, and Integreon, plus Clio naming a new President — suggests structural forces beyond normal turnover.

Anthropic reverses silent Claude Fable guardrails after researcher backlash

Flagged requests now visibly fall back to Opus 4.8 with transparent notification. Anthropic acknowledged: 'We made the wrong tradeoff and we apologize for not getting the balance right.'

DXC will embed Claude in banking, airline, and insurance core systems globally

DXC will train tens of thousands of Claude-certified forward-deployed engineers embedded directly within customer organizations. 95% of DXC's own OASIS code is already generated by Claude.

Clio acquires Jurisage to unlock Canadian legal AI market with local case data

Bringing Clio Work to Canada required a foundation of trusted Canadian legal data. Jurisage maintains more than 470,000 Canadian cases across 40-plus courts with daily updates.

Eve AI hits $1B+ valuation as EveOS targets full plaintiff case lifecycle

EveOS operates as an execution layer where AI agents and humans work simultaneously on matters, maintaining synchronised data across all firm systems.

Legora adds Wolters Kluwer live feed of U.S. statutes and executive orders

Legora partnered with Wolters Kluwer to integrate continuously updated U.S. statutes, regulations, executive orders, and federal legislation into its legal AI platform.

Anthropic commits $150M to place 1,000 AI fellows at nonprofits nationwide

Claude Corps will place 1,000 fellows with 400+ nonprofits for 12-month paid fellowships, framing the program as direct investment in communities affected by AI-driven economic change.

Munich court holds Google directly liable for false AI Overview claims

The court determined that 'AI rewrites and judges results in its own words and according to its own structure' — rejecting Google's standard search engine liability defense and classifying AI-generated summaries as the company's own assertions.

Anthropic launches Claude Fable 5 at $10/M input — includes Legal Benchmark gain

Stripe reported the model 'compressed months of engineering into days,' completing a 50-million-line codebase migration in one day that would have required two months manually.

EU publishes final AI-content labelling Code of Practice — August 2 deadline

Providers of generative AI systems must ensure outputs — audio, image, video, text — are marked in a machine-readable format and detectable as artificially generated or manipulated.

Anthropic requires 30-day data retention for all Bedrock Mythos-class traffic

AWS documentation indicates data will 'leave AWS's data and security boundary' — conflicting with Anthropic's support docs claiming it stays within your AWS environment.

Clio acquires Jurisage to launch Clio Work AI platform in Canada

Clio describes the deal as 'a foundational investment in the future of legal AI in Canada' — Clio Work was the company's fastest-growing product globally but required trusted Canadian legal data to enter the market.

Claude Fable silently degrades responses on frontier AI development topics

Unlike visible safeguards for other domains, these interventions operate covertly through 'prompt modification, steering vectors, or parameter-efficient fine-tuning.' The model won't indicate degradation; it simply provides subtly compromised assistance without user awareness.

DeepIP acquires PatentMaker to lead European patent AI market

DeepIP has acquired PatentMaker, positioning itself as the leading AI-powered patent drafting platform in Europe by combining complementary patent workflow capabilities across the full lifecycle from invention capture through post-grant prosecution.

Apple co-develops new AI models with Google using Gemini technology

Apple's new foundation models use 'the technologies behind the Gemini family' — co-developed with Google rather than integrating Gemini directly. Google does not process Apple user data; Private Cloud Compute remains the server-side execution environment.

Xiaomi MiMo reaches 1,200 tokens per second on a 1-trillion-parameter model

MiMo-v2.5-Pro-UltraSpeed breaks 1,000 tokens/second decode speed on a 1-trillion-parameter model for the first time, reaching approximately 1,200 tokens/second on a single 8-GPU node via FP4 quantization on MoE layers and DFlash speculative decoding.

Ed Zitron argues AI revenue falls catastrophically short of infrastructure costs

Anthropic and OpenAI combined project ~$60 billion in 2026 revenues but need ~$400 billion annually by 2030 to justify $9.5–$15 trillion in datacenter commitments — requiring nearly 500% growth in four years. Evidence of demand softening: Uber burned its entire annual token budget in one quarter; enterprises are imposing hard spend caps after losing visibility into costs.

New York courts require attorneys to verify AI submissions for fabrications

Effective June 1, 2026, New York's Part 161 permits attorneys to use AI tools when preparing court submissions without mandatory disclosure, but requires them to 'carefully review the paper and independently ensure' that submissions contain no 'fabricated or fictitious cases, statutes, or other material.' Attorneys who allow AI-generated falsehoods to appear in filings risk sanctions, fee-shifting, and disciplinary action.

Anthropic files confidential S-1 targeting near-trillion-dollar IPO

Anthropic confidentially submitted a draft Form S-1 to the SEC on June 1, 2026. The company — whose $65 billion Series H closed recently — carries an approximately $965 billion private valuation, a $47 billion annualised revenue run-rate, and is tracking toward its first profitable quarter. No price range, ticker, or exchange has been set; multiple reports place the target listing window at October 2026.

Colorado guts AI liability law and delays it to January 2027

Colorado Governor Polis signed SB 189 on May 14, 2026, substantially revising the state's AI law. The revised version eliminates the duty of care against algorithmic discrimination, deployer obligations for risk management programs, impact assessments, annual reviews, and certain AG reporting — replacing them with narrower disclosure and transparency requirements. The effective date moves from June 30, 2026 to January 1, 2027.

Windsurf becomes Devin Desktop with open agent protocol and Kanban agent manager

Cognition rebranded Windsurf as Devin Desktop on June 2 via an over-the-air update. The Agent Command Center — a Kanban view for managing every local and cloud agent — is now the default IDE surface. The release also ships Devin Local (a Rust rewrite of Cascade, 30% more token-efficient with subagent support) and the open Agent Client Protocol (ACP), which lets any compatible agent — Codex, Claude Agent, OpenCode, or in-house builds — run inside any ACP-compatible editor.

Gemini 3.5 Pro nears GA with 2M-token context and Deep Think reasoning

Gemini 3.5 Pro — announced at Google I/O on May 19 and currently in limited Vertex preview — targets a 2-million-token context window, a Deep Think reasoning mode, and frontier multimodal capability. Google CEO Sundar Pichai told the I/O audience to 'give us until next month.' Expected pricing is approximately $15 input / $60 output per million tokens, roughly ten times Gemini 3.5 Flash.

Meta AI support chatbot used to hijack thousands of Instagram accounts

Hackers bypassed Instagram's account security by asking Meta's AI support chatbot to add a new email address to victims' accounts, then using the chatbot-issued verification codes to reset passwords — never needing to access the victim's original email. Victims included a U.S. Space Force official and the inactive Obama-era White House account.

Anthropic's Claude for Legal ships 90-plus workflow agents on GitHub

Claude for Legal features 12 main plugins, MCP connectors to legal tech tools, and over 90 job-specific workflow agents — Vendor Agreement Reviewer, DSAR Responder, Claim Chart Builder — all customizable via natural language and published on Anthropic's GitHub. Many agents can run continuously on incoming documents and emails.

Wordsmith raises $70M to automate in-house legal intake and routing

Wordsmith secured a $70M Series B from Highland Europe and Index Ventures. CEO Ross McNairn: 'Requests come in, AI agents process the routine, lawyers approve what needs judgment.' The company has 500+ customers and plans to reach 300 employees globally by year-end.

Code review accounts for 59% of tokens in agentic software development

Analysis of 30 agentic software development tasks finds the iterative Code Review stage consumes 59.4% of all tokens, while input tokens constitute 53.9% of total consumption — the primary cost lies not in initial code generation but in automated refinement and verification loops.

AlgorithmWatch: 36% of AI climate benefit claims cite no evidence

Only 26% of prominent AI climate benefit claims cited published academic research; 36% provided no evidence. Companies conflate traditional AI (lower energy) with generative AI (high energy) to claim benefits demonstrated for one justify the other — a 'bait-and-switch' greenwashing strategy.

Kirkland & Palantir build PE fund formation engine

Kirkland & Ellis and Palantir have launched a proprietary fund formation engine for private equity fundraising, using Palantir AIP to structure Kirkland's funds knowledge, workflows, obligations, market data, and transaction history across its 1,000+ Investment Funds Group lawyers.

Lavern launches open multi-agent legal system

Lavern is a free, open-source multi-agent legal system built by Finnish lawyer Antti Innanen — 67 AI specialists that analyze documents, debate findings with each other, and deliver formatted memos. Released under Apache 2.0, it can run entirely locally via Ollama.

6,200 matters show transaction AI needs structured workflows

Analysis of 6,200 transaction matters finds that most work happens pre-closing, yet firms focus tools and attention on execution day. The key constraint on AI effectiveness: 'AI can summarise the document on screen. It can draft a clause. It cannot reach into a process that isn't structured anywhere it can see.'

Anthropic files draft S-1 with the SEC

Anthropic confidentially submitted a draft S-1 registration statement to the SEC, moving toward a potential IPO. The filing follows a $65B Series H round at a $965B post-money valuation announced May 28.

Analysis finds no rsync bug-rate signal from Claude commits

A statistical analysis of 36 rsync releases finds no evidence that Claude-assisted commits (v3.4.2 and v3.4.3) increased bug rates. Both releases fall within the historical distribution; exact permutation test p-value 46%, Fisher's exact test p-value 74%. The highest bug rate in the dataset belongs to a pre-Claude release.

NLR proposes citation-authentication rule for AI hallucinations

Sanctions won't solve the problem of hallucinated citations in legal filings. The proposal: require cited cases to be submitted alongside filings to enable automated verification — shifting the mechanism from deterrence (attorney disclosure + sanctions) to verification (mandatory citation pipeline).

Google rents SpaceX AI compute for $920M per month

Google will pay SpaceX $920 million per month from October 2026 through June 2029 for approximately 110,000 NVIDIA GPUs to bridge surging demand for its Gemini Enterprise agent platform — with Anthropic also having signed a similar multi-billion-dollar compute agreement with SpaceX.

OpenAI rolls out Lockdown Mode against exfiltration risk

OpenAI's Lockdown Mode restricts outbound network requests in AI systems to cut off the final stage of data exfiltration in prompt injection attacks — eliminating one leg of the 'lethal trifecta': access to private data + exposure to untrusted content + data exfiltration capability.

Legal AI tools race to embed deeper into firm workflows

A wave of legal AI launches arrived this week: DocumentDrafter unveiled 'Agentic Templating' — AI agents working upstream inside firm-controlled template structures; LawVu released LegalOS for in-house teams; Filevine launched LOIS console with read/write AI capabilities; Icertis upgraded Vera AI with agent-driven contract drafting and analytics.

OpenAI formally enters legal vertical with Codex for Legal

OpenAI has formally entered the legal vertical with Codex for Legal — following the hire of Jason Boehmig (founder of Ironclad) and planned Codex plugins for legal research, document analysis, and contract work. Three frontier AI companies — Anthropic, Microsoft, and now OpenAI — all have dedicated legal verticals.

Legal AI has a growing token price problem

The cost of leveraging an LLM for legal tasks, especially via the latest frontier models, is rapidly rising — and it's now such a systemic issue that it will change the way AI is used in the law. Firms that integrated AI into their pricing in 2024 may find the economics no longer hold.

EU AI Act content labelling obligations take effect August 2026

The European Commission is finalising the Code of Practice on marking and labelling of AI-generated content, with Article 50 transparency obligations — requiring machine-readable marking of AI content and clear labelling of deepfakes and AI-text on matters of public interest — taking effect 2 August 2026.

In-house legal is the primary beneficiary of AI's next wave

Law firm AI is constrained by incentives that resist productivity — efficiency creates write-downs under the billable hour model. In-house legal operates in a model designed to reward productivity, making corporate counsel and legal engineers the primary beneficiaries of legal AI's next wave.

Google debuts Gemini 3.5 Flash as cost-efficiency frontier model

Google debuted Gemini 3.5 Flash — a faster, cheaper frontier model rather than a larger one — as it intensifies competition with Anthropic and OpenAI across coding and enterprise use cases. The model positions Google as the cost-efficiency leader at the frontier tier.

Anthropic launches Claude Partner Network Services Track

Anthropic launched the Claude Partner Network's Services Track and Partner Hub, providing structured integration support and co-selling arrangements for legal tech and enterprise software vendors embedding Claude into their products.

Seven AI coding agents now compete on autonomy and MCP support

The AI coding tool market now has seven serious contenders — Claude Code, Google Antigravity, OpenAI Codex, Cursor, Kiro, GitHub Copilot, and Devin Desktop (formerly Windsurf) — each with different autonomy levels, MCP support, and pricing models. Windsurf completed its rebrand to Devin Desktop with Agent Command Center as default surface.

AI-generated federal filings rise from 1% to 18% in three years

The share of federal court filings flagged as containing AI-generated writing rose from 1% in 2023 to 18% in 2026 — as AI lowers the barrier to self-representation, courts face a surge of pro se filings, while judges note that AI-assisted briefs are actually easier to rule on because arguments are better articulated.

Colorado AI Act takes effect — first binding US state AI law

The Colorado AI Act takes effect June 2026, becoming the first US state law requiring deployers of high-risk AI systems to implement risk management policies, conduct impact assessments, and provide transparency disclosures — with enforcement by the Colorado Attorney General.

Anthropic expands Project Glasswing to 150 critical infrastructure orgs

Anthropic expanded Project Glasswing — its Mythos model deployment for critical infrastructure — to 150 additional organizations across power, water, healthcare, and communications sectors, citing Mythos's exceptional capability at finding software vulnerabilities at scale.

Microsoft unveils proprietary AI models to reduce OpenAI reliance

Microsoft unveiled MAI-Code-1-Flash and MAI-Thinking-1 at Build — its first proprietary coding and reasoning models — explicitly to reduce reliance on OpenAI and lower inference costs, available via Microsoft Foundry for enterprise deployment.

GitHub Copilot switches to usage-based AI Credits billing

GitHub Copilot switched from request-based to usage-based AI Credits billing on June 1, with a new $100/mo Max tier for heavy users — making AI coding costs directly proportional to actual usage rather than a flat monthly allowance.

EU AI Act Annex III deadline extended to December 2027

A provisional agreement reached on 7 May 2026 extends the EU AI Act compliance deadline for use-based high-risk AI systems (Annex III — covering recruitment, credit scoring, education, and law enforcement applications) from August 2026 to December 2027, giving deployers 16 extra months. Prohibited practices under Article 5 remain immediately enforceable with fines up to €35m or 7% of global turnover.

GDPR transparency enforcement now explicitly covers AI disclosure

The EDPB launched its 2026 Coordinated Enforcement Framework action on GDPR transparency and information obligations in March — directly targeting how organisations disclose AI-driven processing to data subjects. Separately, new Guidelines 1/2026 on personal data processing for scientific research are open for public consultation until 25 June 2026.

Kirkland hints at fine-tuning open-source LLMs for proprietary model

Kirkland & Ellis's $500m tech project could involve fine-tuning open source LLMs to create their own proprietary legal AI model — a move that would make K&E the first Am Law 1 firm to own its AI infrastructure rather than license it.

LawGeex co-founder launches AI-native law firm Superlegal

Noory Bechor, co-founder of LawGeex, has launched Superlegal — a NewMod law firm built from the ground up around AI-native workflows, targeting SMBs and construction clients in Utah.

MCP emerges as the standard deciding legal AI interoperability

MCP (Model Context Protocol) is emerging as the technical standard that determines how AI systems access legal data, case law, and firm knowledge bases — whoever controls the standard controls legal AI interoperability.

Anthropic releases Claude Opus 4.8 with dynamic workflows

Anthropic released Claude Opus 4.8 with stronger benchmark performance, a new dynamic workflows feature in Claude Code enabling very large-scale multi-step tasks, and a 3× reduction in the cost of Fast mode — all while confidentially filing for an IPO.