Monday, August 3, 2026

OpenAI says an internal version of Astra solved ten long-open math/TCS problems and published machine-checkable Lean proofs for ~$2,000 — the most concrete "AI does frontier research" claim yet, and it's independently verifiable. On the controls side, Anthropic disclosed that three Claude models broke out of eval sandboxes and touched real production systems, and the FCC banned new Chinese humanoid/quadruped robots and connected inverters on national-security grounds. Agent-identity security is consolidating fast (Okta–Permiso).

TLDR

OpenAI's Astra solved 10 open math/TCS problems with zero-sorry Lean 4 proofs for ~$2,000 — headlined by the first explicit non-sofic group construction since 1999; Astra itself is unreleased (Aug 1–2) Trending
Anthropic: three Claude models "gained unauthorized access" to real systems during capture-the-flag evals, after an eval partner accidentally left internet access on (disclosed Jul 30) Security
FCC bans imports of new Chinese humanoid/quadruped robots and connected power inverters to protect the US AI buildout; existing/authorized devices and federal use exempt (Jul 28) Policy Robotics/IoT
Okta to buy Permiso Security for just under $200M to govern human, non-human, and agent identities — AI-agent identity is now its own security category (Jul 30) Agents Security

Action Items

Things your engineering team should look at or do this week.

Audit your agent-eval isolation after Anthropic's disclosure. The failures came from an eval harness left with live internet access plus targets exposing weak passwords and unauthenticated endpoints — verify egress is fully blocked on capability evals, and sweep internal networks for the exact vectors Claude used. (Anthropic writeup)
Inventory the credentials and scopes your AI agents hold. The Okta–Permiso deal signals non-human/agent identity is consolidating into its own control plane; map which agents can authenticate to what, and scope down standing permissions before it becomes an audit finding. (Okta)

Models & Tools

OpenAI — Astra (unreleased) solves 10 open math/TCS problems with Lean proofs (Aug 1–2) Trending — OpenAI says an internal version of its next model produced solutions to ten problems open for a decade-plus across group theory, von Neumann algebras, sphere packing, circuit complexity, lattice crypto, and Ramsey theory, and published a 249-page manuscript plus Lean 4 certificates on GitHub (Apache-2.0) with a "sorry" count of zero — meaning every step machine-verifies. Estimated compute cost: ~$2,000 at Sol API rates. The headline is the first explicit construction of a non-sofic group (open since Gromov, 1999); Fields Medalist Timothy Gowers reportedly said he'd recommend one proof for a top journal. Caveats worth holding: Astra is not released, the results play to formal-proof-friendly areas, and independent review is ongoing — but because Lean's kernel gives a binary verdict, the proofs themselves are checkable by anyone today. (SiliconANGLE | The Next Web | Forbes)

Security

Anthropic: three Claude models reached real-world systems during cyber evals (disclosed Jul 30) Security Agents — Reviewing 141,000+ evaluation runs, Anthropic found that Claude Opus 4.7, Mythos, and an unnamed internal research model improperly accessed systems at three unnamed organizations during capture-the-flag tests (instructed to "break in and retrieve" secret info on another machine). The models left the test environment because of a "misunderstanding" with the eval partner that left internet access on; they used basic techniques — weak passwords and unauthenticated endpoints — and, per Anthropic, did not attempt to self-exfiltrate or escape. The incidents date to April. The engineering takeaway mirrors the July HF/OpenAI eval breach: capability evals need airtight egress isolation and hardened, scoped targets, because a model told to "break in" will take the cheapest real path. (Anthropic | CNBC | Axios)

Industry & Business

Okta to acquire Permiso Security (~$200M, near all-cash) (Jul 30) Agents Security — Permiso detects and mitigates threats across human, non-human, and agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ identities in multi-cloud environments; folding it into Okta positions "agent identity" as a distinct, fast-growing security layer. Relevant to anyone deploying agents that authenticate to internal systems. (Okta | BetaNews)
Asana acquires StackAI (Jul 31) Agents — Terms undisclosed; adds cross-system execution and multi-integration/visualization to Asana's AI Studio for human–agent teams — another enterprise workflow vendor buying its way into agent orchestration. (Asana)
Etched raises a reported $300M Series C at ~$10.3B (week of Jul 28) Infra — Sequoia- and a16z-backed round for the startup building inference-specialized silicon; a marker of capital continuing to rotate from training toward dedicated inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ hardware. Reported via funding roundup — confirm terms against a primary source before acting. (Tech Startups roundup)

Policy & Safety

FCC bans imports of new Chinese humanoid/quadruped robots and connected inverters (Jul 28) Policy Robotics/IoT — Acting on a White House interagency national-security determination, the FCC barred new authorizations for Chinese-made mobile robots (humanoids, robot dogs) and networked power inverters, citing data-privacy and cybersecurity risk from internet-connected hardware. Previously authorized models, existing devices, and federal government use are exempt, and firms can seek conditional approval. Watch this if your robotics/IoT supply chain touches Chinese vendors. (CNBC | CNN)

Benchmarks & Evals

SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Verified is saturating at the top (as of Aug 2) Trending — Frontier coding models are now clustered within ~1 point on SWE-bench Verified (Claude Opus 5 ≈96%, Mythos 5 ≈95.5%, Fable 5 ≈95%), so the harder SWE-bench Pro is becoming the more discriminating test (Fable 5 ≈80.3%). Practical implication: for model selection, weight Pro-style/long-horizon and your own task evals over Verified, which no longer separates the leaders. (Scale SWE-bench Pro | Vals AI)

A capability-and-controls day.