← /editions
#65

The hacking model has a guest list

// OpenAI ships GPT-5.6-Cyber, a model trained not to refuse exploit-chain work, gated by vetting rather than price — while the cost of an ordinary agent turn collapses again and agents start carrying corporate cards.

$ today_in_10_seconds

OpenAI is now shipping a model trained specifically not to refuse exploit-chain work — to a vetted list, under legal attestation. The same week, the cost of an agent turn fell again and startups began issuing corporate cards to agents. Capability is being unbundled from permission, and the unbundling runs in both directions.

$ lead_story

OpenAI ships an offensive-security model behind a guest list

On Monday OpenAI split its Daybreak cyber program into two tiers and introduced GPT-5.6-Cyber. Daybreak Blue gives approved defenders GPT-5.6 Sol with the system-level safeguards that normally screen security requests removed; Daybreak Red unlocks GPT-5.6-Cyber, a Sol variant trained for zero-day discovery and exploit-chain development. The gap is the story: on OpenAI's internal advanced-cybersecurity benchmark — exploit chains, authentication bypass, privilege escalation — GPT-5.6-Cyber answered 95% of requests, against 1.5% for standard Sol and 2% for the Blue-tier model. OpenAI used it against V8 and found two chainable memory-corruption bugs capable of escaping the heap sandbox, now tracked as CVE-2026-15903. Access runs through identity verification, monitoring, approved-use restrictions and legal attestations; press reporting names Accenture, IBM, CrowdStrike and Cloudflare among the first partners, which OpenAI has not confirmed. Read the sequencing honestly. This lands days after OpenAI delayed Astra for reaching critical hacking capability in testing, and while it is still investigating how its own tools breached Hugging Face — GPT-5.6-Cyber only hit 'High' on the Preparedness Framework's cyber threshold, not Critical. So the argument is that defenders need this before attackers field it at scale, made by a company that has spent three weeks demonstrating that lab containment is not reliable.

// why_it_matters: The safety boundary just moved out of the model and into a vetting list — which means someone else's access-review process is now part of your threat model.

read more →

$ also_worth_attention

  • The builder's guide to GPT-5.6

    Yesterday's guide is the clearest artifact yet of how fast agent economics are moving. On BrowseComp, GPT-5.5 at Extra High scored 84.36% for $33.27 three months ago; GPT-5.6 Luna now scores 84.04% for $1.33 — and prices have dropped again since. On Agents' Last Exam, Sol at low reasoning effort beat GPT-5.5 at high, harness held constant. Treat the head-to-head claims against Fable 5 as vendor benchmarks; the independent read is narrower — Claude Opus 5 is roughly Sol-class and about 6% cheaper. The actionable part isn't the pricing, it's that your reasoning-effort settings are now miscalibrated: prompts and effort levels tuned on 5.5 are costing you accuracy, not buying it.

    link →
  • Nvidia opens Switchyard, and routing stops being a hack

    Nvidia shipped Nemotron 3.5 Lightning on Tuesday alongside NeMo Switchyard, an open-source library that routes each request in an agent workflow to the best-fit model across your own mix of open, proprietary and Nvidia models without rewriting the application. Nvidia's internal benchmarks claim frontier-level accuracy at roughly a third of the cost of running Opus 4.8 alone, but the useful number is LangChain's own: across 145 multi-turn Deep Agents tasks, only 7% of turns actually needed a frontier model, and routing the rest to a 30B cut total cost 74% for six points of accuracy. Six points is a real price. And the router is a new failure surface — bad classifications, stale cost data and botched handoffs can eat the entire saving, so evaluate on final agent outcomes rather than per-call quality.

    link →
  • Mercury gives agents their own corporate cards

    Also Tuesday: Mercury launched Spend, with budgets, self-enforcing expense policies and a card class built for AI agents. A human has to create the card; the agent spends inside its limits and cannot change them, and the card stays trackable, auditable and cancellable, with over-budget transactions declined at the point of sale and merchants restricted by name or MCC. Mercury's Akhund says customers were already hand-issuing virtual cards to agents — which is the real signal. Note what the control model actually is: not approval-per-action, but a pre-authorized envelope the agent operates inside. That's the same pattern as Daybreak, one layer down the stack.

    link →

$ bigger_picture

The trend behind the headlines

Three of today's stories drive the marginal cost of an agent action toward zero: $1.33 buys what $33 bought in May, 93% of agent turns don't need a frontier model, and an agent can now hold a payment credential without a human in the loop. The fourth does the opposite — one capability so consequential that OpenAI won't sell it per token at all, only per vetted relationship. That's the shape of the next two years: almost everything an agent does becomes too cheap to meter, while a narrow band gets rationed by identity instead of price. Both halves hand you the same bill. When capability is gated by who you are rather than what you pay, and spend is gated by a budget envelope rather than an approval, your attestation and audit trail stop being compliance artifacts and become the control plane. Most teams have neither wired to an agent today.

$ tool_of_the_day

LiteLLM AI Gateway

If you're going to route across models and hand agents their own budgets, do it at one egress point rather than in application code. LiteLLM is a self-hosted gateway with virtual keys, cost tracking and an admin UI, exposing 100+ providers behind one OpenAI-compatible API with per-key budgets and rate limits. The underrated feature for agent work is budget fallbacks: when a key exhausts its cap on one model, the request is rerouted to the first fallback still in budget rather than failing. Know the constraint before you deploy — budgets and virtual keys need Postgres, since SQLite races on budget decrements under concurrency.

explore LiteLLM AI Gateway →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime