$250K pre-seed on a post-money SAFE with a $4M valuation cap. You own 6.25% when it converts. Optionally, stake through Unblocks and earn a share of $SAFEBUX cashflow while the network grows — income from real usage, not a promise about a future exit. The position appreciates as consumer adoption drives the valuation from $4M to $10M at the January 2027 VC round and toward $50M at the Series A round that is expected to follow.
The product is built and live. Three clients are already running on Safebots: unaligned.club (Scoble's AI community), caroline.vote (Congressional campaign), freecities.app (international conferences). We compiled 49 tools for ten cents. The economics are measured. The $250K starts the go-to-market that turns 1 million existing community-leader emails into the hockey-stick growth VCs price the January round on.
Dates are when each raise opens, not when it closes. These are targets, not commitments.
U.S. accredited investors come in under Regulation D. Eligible non-U.S. investors can take the same economics offshore under Regulation S, in a form that can later be tokenized and traded. Both fund the same balance sheet.
The straightforward private-placement route for U.S. accredited investors.
Same economics, structured so eligible non-U.S. investors can hold and trade a tokenized security.
Every time a Safebot runs — every document processed, every workflow executed, every hour of compute — the customer pays in $SAFEBUX. A portion of $SAFEBUX sale proceeds flows to $SAFE stakers, pro-rata. The yield tracks actual usage, the way Ethereum staking yield tracks gas fees. More users, more operators, more Safebots running — more cashflow to stakers.
Without staking: you put in $250K, own 6.25% on conversion, and wait for appreciation. With staking: you earn a share of network revenue from day one while the position appreciates underneath.
Every company in the comparable set repriced faster than this. Three rounds, dated by when each raise opens.
| Round | Opens | Raising | Valuation | Share |
|---|---|---|---|---|
| Pre-seed | Sep 2026 | $250K | $4M cap | 6.25% |
| VC round | Jan 2027 | $2M | $10M | 20% |
| Series A | Feb 2028 | $15M | $50M | 30% |
Entry price. The cheapest the company will ever be.
1.0× — cost basisSAFE converts at the cap, then dilutes from 6.25% to ~5%. Company worth 2.5× more.
~$500K on paper — 2×~3.5% of $50M. Seventeen months after the pre-seed.
~$1.75M on paper — 7×We compiled the fundraising history of every AI infrastructure and agent-safety company that raised a seed or Series A since 2023. Safebots enters at $250K at a $4M cap — well below the mean — on a product that's already live.
| Metric | Seed | Series A |
|---|---|---|
| Median raise | $4.5M | $20M |
| Mean raise | $5.5M | $26M |
| Median valuation | $25M | $120M |
| Median time seed→A | — | 12 months |
| Safebots (this raise) | $250K at $4M | $15M at $50M (target) |
Safebots enters below the median on both raise and valuation. The entry price is deliberately conservative — which means the repricing at the VC round and Series A has more room to run.
AI safety / governance: Anthropic ($750M seed, $4.1B A), Inflection ($225M seed), Adept ($350M A), Cohere ($270M A). These are the frontier-lab-scale raises. Below them: Robust Intelligence ($30M A), Lakera ($20M A), Calypso AI ($14M seed).
AI infrastructure: Together AI ($20M seed, $102M A), Anyscale ($40M A), Modal ($16M seed), Baseten ($20M seed, $40M A), Fireworks ($25M seed). These companies sell inference or compute — the same layer Safebots occupies, from the other side.
Agent platforms: Fixie ($17M seed), Relevance AI ($15M seed), Dust ($5M seed), CrewAI ($18M A). These are the companies Safebots competes with directly — and whose architecture (open-ended agents) is the problem Safebox solves.
OpenAI launched agents in sandboxes. They found a shared channel, coordinated for days, hacked Hugging Face, and spoofed their own transcripts. METR investigated and could not rule out that AI agents misrepresented the findings.
PocketOS (Apr 2026): A Cursor+Claude agent found a token and deleted the Railway production database in nine seconds, including backups.
DataTalks.club (Feb 2026): Claude Code ran terraform destroy with auto-approve. Two and a half years of student data wiped.
SaaStr/Replit (Jul 2025): Told in capital letters eleven times not to touch production. Deleted 1,200+ contacts, fabricated 4,000 records, and said rollback was impossible.
Opus 4.7 (Apr 2026): Mass-emailed the entire customer database up to twenty times per contact. The safety rule was in CLAUDE.md. The model read it.
On August 27, 2026, OpenAI, Anthropic, AWS, Microsoft and 100+ other companies signed an open letter warning that organizations have months, not years to prepare for AI-enabled cyberattacks. They named AI-generated code as something to demand more of. That is a procurement trigger with a date on it — and no commitments, deadlines, or funding attached.
Frontier labs spend billions moving the model from 9 to 9.5. The infrastructure around it (declared workflows, compiled tools, local models, domain knowledge) sits at 2. The factors multiply, so raising it to 4 doubles capability on the same model. Safebots builds that layer. The scores are illustrative, not measured.
The model is near its ceiling and the infrastructure is barely built. Each model gain costs more than the last.
Same model with twice the infrastructure: a 100% improvement, at a fraction of the cost of a better model.
A Safebox is not a chatbot. It is not an agent. It is a contained environment — running on your infrastructure or ours — where compiled workflows run under four rules the model cannot override: declared workflow steps, signed limits on each tool, an approval gate on every side effect, and an audit trail the model cannot write.
Safebox is a single isolated environment. One organization, one set of workflows, one audit trail. Runs on any cloud or on-premises hardware. This is the unit of deployment.
Safenet is the network of Safeboxes. One organization can grant another access to specific tools, with a cost cap, an expiry, and a rate limit. Organizations can also share tools and negotiate agreements among several parties.
Safecloud is the managed service. We run the Safeboxes, manage the models, handle updates. The customer gets an endpoint and a dashboard. For organizations that want the economics without the ops.
We gave an AI a $4 budget and said: build tools for a law firm, a hospital, an online store, a university, a news company, a real estate agency, a startup, a nonprofit, a consulting firm, and a tech company. It built 49 tools and had $3.90 left over.
| If each tool runs… | Safebots | Agent swarm | Savings |
|---|---|---|---|
| once | $0.044 | 10¢ | 2× |
| 10 times | $0.044 | $1.00 | 23× |
| 100 times | $0.044 | $10 | 227× |
| 1,000 times | $0.044 | $100 | 2,273× |
The Safebots line is flat. The agent line grows linearly. Model: Claude Haiku 4.5 (cheapest available). Every number measured, not projected. Full session log →
Pure compute ($0/run): Regex, FFmpeg, math, validation, format conversion. 42 of top 100 workflows.
Local model ($0/run): Classification, sentiment, PII detection, language ID. 3B param model in WASM. 26 of top 100.
Cloud LLM ($0.001–0.01/run): Complex drafting, multi-step reasoning. 32 of top 100. Used only when local capability is exceeded.
We manage the open-source models (Llama, Phi, Mistral, Whisper) — quantized, tested, deployed inside the Safebox. No MLOps team required. See all 100 workflows with costs →
Most of the industry treats autonomous inference as the product. Safebox treats inference as a temporary method for discovering dependable computation. An agent explores, finds a trajectory that works, and that trajectory gets compiled into a workflow. From then on, the task executes deterministically at zero LLM cost.
An agent finds a working sequence once; after that a compiled workflow runs it without the agent. A conventional agent pays the reasoning cost every time. Safebox learns a successful path once, encodes it, and reruns it without calling the model. Each cycle widens the cost gap.
Developers left Ethereum for Polygon and Solana because the same transaction cost far less there. Safebots makes the same move in AI: same outcome, 1000× cheaper execution. The blockchain parallel → · Full economic case →
An April 2026 study from Stanford, MIT, Michigan, DeepMind, and Microsoft AI measured agent token consumption across 500 tasks. Input-to-output ratio: 153:1. Cost variance on the same task: up to 30×. Higher token consumption does not correlate with higher accuracy.
Safebox attacks both sides: fewer tokens per outcome (workflow architecture) and lower cost per token (self-hosted open weights). Even if API prices fall, the workflow efficiency advantage persists. Read the paper →
Every other startup wraps an open-ended agent in guardrails that detect damage after it happens. In a Safebox the model has no direct access to production. It runs only declared steps with signed tools, outside changes need M-of-N approval, and the box, not the model, writes the audit log.
Monitors, filters, review steps layered on a system that can still reach production. Detection is an admission that the action was possible.
Declared workflows, signed tool limits, an approval gate on every side effect, and an audit trail the model cannot write.
EU AI Act: in force. High-risk AI systems require documented risk management, data governance, and human oversight.
SEC AI disclosure: rules are live. Public companies must disclose AI-related risks.
HIPAA: BAAs now gate AI procurement in healthcare. No BAA, no deployment.
Gartner: expects 75% of enterprise AI workloads to require attested compute by 2029.
A CISO who was going to evaluate agent governance next year now has to answer for it this quarter. Our cost claim gets the meeting. This is what turns the meeting into a purchase.
Frontier labs charge per token. An architecture that touches the model rarely is not one they can ship.
Agent startups are built on an open-ended loop. Structural containment requires that the agent never had the reach — that means rebuilding the product.
Clouds and integrators are the intermediary being removed. Proving the operator can't read your data is a strange thing to buy from the operator.
VCs want two things: a hockey-stick growth curve and recurring revenue. Path 1 produces both from an existing audience. The other three make them bigger and faster.
Produces: hockey-stick growth + recurring revenue — the two things VCs price on
Six million people downloaded the Groups app portfolio. About a million left an email address. 38,000 devices are push-reachable. Five projects already live. The $250K funds the automated sales funnel that converts them into paying weekly customers.
Step 1: Push notification or email. "Your community now has an AI assistant."
Step 2: Try free. Branded app + bots. No credit card. Members start joining.
Step 3: Members build history. Relationships, content, workflows. This is the lock-in.
Step 4: Leader charges members for access. Revenue starts flowing.
Step 5: Free tier ends, weekly billing begins. Cancelling means losing the community.
| Metric | Jan 2027 | Jun 2027 | Sep 2027 | Mar 2028 |
|---|---|---|---|---|
| End users | 2,000 | 25,000 | 200,000 | 1,000,000 |
| Leaders | 80 | 800 | 5,000 | 15,000 |
| Paying | 5 | 60 | 500 | 2,000 |
| MRR | $5K | $40K | $350K | $1.8M |
Each leader brings 50–200 members. That multiplier turns 1,000 leaders into 25,000 users. The Groups app did this seven million times.
YouTube (2–3 videos/week), paid crypto influencer distribution, influencers as customers — we build them their own membership site. Scoble, Edge, Shinkle already live. Each influencer promotes their own Safebots-powered site to their audience, at no cost to us.
Alex Yampolskiy (CEO) and Sam Kassoumeh (cofounder) invest and mentor. Art Basel Miami December: 30+ CISOs. SecurityScorecard detects AI governance gaps — Safebox is the remediation. $50K–$500K per contract. Warm intros to BoldStart, GV, Sequoia.
Recently departed frontier-lab executive who left over agent safety concerns. Scoble makes introductions. This person pitches Sequoia, a16z and YC as a lab executive who left over agent safety and joined the team addressing it. They open the $20M round.
Path 1 alone produces the growth curve. Path 2 amplifies it. Path 3 anchors revenue. Path 4 opens institutional doors. A startup that depends on one path is fragile. Full roadmap →
All three are live. Each runs in its own Safebox, so members and data stay inside that community. Each brings a different audience onto the same platform.
AI community. Founders, technologists, investors, researchers. Has interviewed Musk and Zuckerberg. His audience is the CEO pipeline for Path 4. His members use the site for a member directory, video clips and events.
Civic / campaign2026 Republican nominee, NY-12. Manhattan professionals, donors, civic groups, volunteers. Real-world testbed for outreach, events, constituent follow-up. The site runs an issue quiz, testimonial recording, events, chats and a supporter directory.
Conference / internationalGlobal gathering around new governance models. Founders, investors, crypto holders, international organizers. Conference → community → dealflow. Attendees use it to see the schedule and speakers and to find others with shared interests.
Branded app + AI bots for their community. Members join, interact, build history. No credit card. This is the acquisition layer.
Lock-in: members and content accumulate. Leaving means losing it.After the free trial, weekly billing starts. $20–200/week depending on usage. Replaces Zapier + ChatGPT + Mailchimp + scheduling tools.
Revenue: recurring, weekly, usage-based. Grows with community size.$50K–$500K/year. Deployed in their VPC. An isolated environment and database per team. Designed for HIPAA and GDPR deployments; not yet SOC 2 certified. SecurityScorecard warm intros.
Revenue: contract, annual, high-margin. Anchors ARR for VC rounds.We catalogued the 100 most popular AI workflows across content, sales, operations, legal, finance, HR, support, data, engineering, and education. Total compilation: eleven cents.
Lead generation. Setting appointments. Sourcing. Sales workflows. Cutting videos, translating them, making thumbnails, generating versions in different languages, adding intros and outros. Working all night on transcribing, cutting, publishing — holding the results until morning for approval. Having a panel of reviewers vote on variants before publishing.
That's not "AI agents." That's compiled workflows doing reliable work at a fraction of a cent per run.
Cut, watermark, transcribe, translate, repurpose, moderate. Most at $0 (FFmpeg, Whisper).
Qualify leads, personalize emails, prep meetings, generate proposals. Local model for scoring.
Summarize inboxes, audit timesheets, compile status reports. Mostly pure compute.
Extract clauses, compare NDAs, redact PII (pure regex), scan regulatory changes.
Parse invoices, reconcile statements, forecast cash flow. All pure compute or local model.
Screen resumes, generate JDs, analyze surveys, calculate PTO. Mixed tiers.
Classify tickets, auto-respond, detect escalation, find KB gaps. Mostly local model.
Clean CSVs, detect anomalies, build retention cohorts, aggregate dashboards. Pure compute.
Plus 8 engineering/DevOps workflows and 5 education/nonprofit workflows. See all 100 with costs →
$250K carries the company to the January 2027 VC round. The core product is built and live — this budget covers the team that operates it, the engineering that hardens it for enterprise, and the distribution that produces the growth curve VCs price the next round on.
$SAFEBUX is the gas token for Safebox compute. Every workflow, every tool run, every hour of uptime costs $SAFEBUX. Customers buy it; stakers earn it. The yield tracks real usage.
Cashflow from real network usage starts in the first months, before any equity event.
$SAFEBUX listed for trading. Liquid income stream with a visible market.
Company-approved partial sale at ~$50M. Return the $250K while most of the position keeps riding.
IPO, M&A, or secondary sale at Series B/C prices. The final realization on the residual position.
Safebots, Inc. is a Delaware C-Corp formed via Stripe Atlas (May 6, 2026). The SAFE issues no shares today. It converts into preferred stock at the next priced equity round, at the $4M post-money cap, fixing your ownership at 6.25% before that round's new money.
| Detail | Value |
|---|---|
| Corporation | Safebots, Inc. — Delaware C-Corp |
| Raise | $250,000 |
| Instrument | Post-money SAFE (Y Combinator standard form) |
| Valuation cap | $4,000,000 post-money |
| Your ownership | 6.25% on conversion |
| Conversion | At the next priced equity round (January 2027 target) |
| Token optionality | Deposit the SAFE via Unblockers → $SAFE → stake for $SAFEBUX yield |
| Holder | Shares | % |
|---|---|---|
| Founder (Greg Magarshak) | 8,000,000 | 75.0% |
| SAFE investor | 666,667 | 6.25% |
| Option pool | 2,000,000 | 18.75% |
| Total | 10,666,667 | 100% |
Illustrative: share counts assume conversion at the cap ($0.375/share). Authorized shares: 10,000,000, so conversion will require a board increase. Founder shares vesting 4 years with 1-year cliff per Stripe Atlas standard. COO equity grant details →
