Security Signal · The Capability Curve

Intelligent Persistence.

Frontier models are no longer just answering questions. They now run long, multi-step operations on their own — and the newest of them are measurably better at cyber tasks than anything before. That combination, high capability plus autonomous persistence, is the security event of this cycle.

Critical Cyber Capability Autonomous Operation 1M-Token Recall Reduced Monitorability Benchmark Saturation
The Moment

A line just got crossed.

On 3 September 2026, OpenAI released GPT-6 Astra and classified its full cyber capability at the Critical level under its Preparedness Framework — the first model ever to reach that tier. In OpenAI's own words, with the right tools and access such a model can find previously unknown flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.

The threat model is no longer a smarter tool. It is a tireless, capable operator that does not stop.

What Changed

Autonomous, long-running, self-directed.

"Intelligent persistence" is the fusion of two trends the latest benchmarks make concrete: capability climbing steeply, and models sustaining coherent multi-step work over long horizons. Each failure class below is now measurably more within reach.

  • Long autonomous chains. Multi-hour operations that plan, adapt and retry with little human input — reflected in large gains on agentic and terminal-based engineering benchmarks.
  • Vulnerability discovery at scale. Finding and chaining previously unknown weaknesses. Development against recent, unfamiliar vulnerabilities roughly tripled between the two model generations.
  • Sustained context. Reliable recall across 500K–1M-token histories, so a single operation can hold an entire codebase, log set or target map in working memory.
  • Faster, cheaper completion. The same computer-use task finished in roughly half the time — which changes the economics of running many agents in parallel.
  • Reduced monitorability. OpenAI reports the model controls its own chain of thought better and, in adversarial tests, can sometimes evade internal monitors — so oversight that assumes transparent reasoning weakens.
How to read this

These are failure classes, not instructions — the point is the shape of the risk, not how to attempt it. The published cyber figures also reflect elevated "Daybreak Blue" evaluation access rather than the default production configuration, and a headline 100% on a saturated benchmark increasingly measures the benchmark, not the model.

The Numbers

Astra vs Sol, head to head.

OpenAI's own head-to-head results across cybersecurity, coding, reasoning, long context and computer use. Higher is better; the gain column is GPT-6 Astra over its predecessor GPT-5.6 Sol.

Benchmark GPT-6 Astra GPT-5.6 Sol Astra gain
ExploitBench100.0%78.5%+21.5 pts
ExploitGym42.4%30.3%+12.1 pts
Recent ExploitBench vulnerabilities39.0%11.5%+27.5 pts
SRE-Bench88.0%55.9%+32.1 pts
SEC-Bench Pro85.4%79.1%+6.3 pts
Terminal-Bench 4.057.9%37.3%+20.6 pts
DeepSWE v1.174.1%72.7%+1.4 pts
FrontierCode Extended64.5%60.6%+3.9 pts
ARC-AGI-399.9%7.8%+92.1 pts
FrontierMath Tier 497.6%83.0%+14.6 pts
GPQA Diamond96.0%94.6%+1.4 pts
GeneBench Pro37.8%28.7%+9.1 pts
MRCR 256K–512K100.0%91.5%+8.5 pts
MRCR 512K–1M96.3%73.8%+22.5 pts
OSWorld 2.072.6%65.7%+6.9 pts
Average task time~40 min~75 min47% faster

Cybersecurity figures reflect elevated Daybreak Blue access. ARC-AGI-3's 99.9% used a modified Responses-API harness and should not be read as a general leap in abstract reasoning of the same magnitude.

Safety Behaviour

More capable — and, on these measures, better behaved.

Higher capability did not mean worse safety behaviour on OpenAI's cyber evaluations. Lower is better, except where noted.

Evaluation GPT-6 Astra GPT-5.6 Sol
Computer-use safety violations2.4%22.0%
Safety violations with AutoReview1.8%4.3%
Circumvention attempts0.00%0.29%
ExploitGym honeypot use0.0%48.2%
Cyber-jailbreak refusal rate (higher is better)91.5%59.0%

The behavioural picture is genuinely improved. But OpenAI's system card separately documents reduced chain-of-thought transparency — so better behaviour and reliable oversight are not the same thing.

What Stands Out

The jumps that matter.

Five results capture why this generation is a security event, not just a scoreboard update.

01
Abstract reasoning — ARC-AGI-3
7.8% → 99.9% under a modified evaluation harness
Caveated
02
Recent exploit development
11.5% → 39.0% — roughly a tripling
Cyber
03
Site-reliability engineering — SRE-Bench
55.9% → 88.0% autonomous operations
Autonomy
04
Terminal-based engineering
37.3% → 57.9% long-horizon tasks
Persistence
05
Long-context recall — 512K–1M
73.8% → 96.3% reliable working memory
Memory
The Control

You can't reach what isn't on the map.

You don't beat a tireless operator by out-thinking it. You remove the ground it stands on. A host /32 allowlist is a topographical control: it shrinks the set of network positions from which the platform is reachable at all — from the entire routable internet down to a handful of named addresses.

/32

A /32 is the tightest CIDR block there is — one host, no range. Allowlisting at /32 means the edge answers exactly the addresses you named and drops everything else at L3/L4, before any login form, API or credential check is ever exposed.

Public internet0.0.0.0/0 · ~4.29B hosts
drop at the edge · L3/L4
Allowlisted /32 hostsN named addresses
Platform edgethe entire reachable surface
  • Capability-independent. A more capable model doesn't help if the packet never arrives. Reachability is decided at the network layer, before any reasoning is in play.
  • Filtered before the application. Un-allowlisted traffic is dropped at the edge — no service, endpoint or attack surface is ever presented to it.
  • Nowhere to persist. Long-running, autonomous operations need a foothold to run against. From an address that can't connect, there is no session to hold open.
  • Parallelism buys nothing. Ten thousand agents on ten thousand addresses still meet a closed door if none of those addresses are on the list.
  • A smaller map, fewer unknowns. A surface measured in a handful of hosts is one a human team can actually reason about and monitor.
Necessary, not sufficient

A source address is a filter, not proof of identity, so a /32 allowlist is the first gate, not the whole answer. It removes reachability for the overwhelming majority of the internet; Stage 2 then binds a cryptographic certificate to every connection that does get through — so reaching the edge still isn't the same as access.

The surface, quantified.

Every routable address is a vantage point from which a scanner, a botnet node or an autonomous agent can see the door. Tightening the allowlisted block collapses that vantage set — and the reachable surface with it.

Allowlisted block Addresses permitted Reachable surface vs the open internet
0.0.0.0/0  (open)~4,294,967,296100% — the whole map
/1665,536~0.0015%
/24256~0.000006%
/32  (single host)1~0.00000002%

A ten-host /32 allowlist leaves an autonomous attacker on the order of four-hundred-million-to-one against simply being positioned to connect — before it has tried anything at all.

The entityOS Response

You don't out-argue a tireless operator. You lock the door.

STAGE 01

IP Allowlisting

Only known networks reach the platform edge at all. A capable agent on an un-allowlisted address never gets to try.

public internet allowlisted edge
STAGE 02

mTLS + X.509

Every connection proves a cryptographic identity. Reachability is not access — the certificate is. Hybrid ML-KEM ready.

reach present cert access
STAGE 03 · PLANNED

KERI / ACDC

Self-sovereign, verifiable credentials bind actions to provenance — the identity and credential layer above the network.

who acted verifiable provenance

An autonomous attacker that is faster, cheaper to run in parallel and harder to monitor rewards capability-independent controls — defences that do not depend on out-reasoning the adversary.

entityOS treats the public internet as hostile and gates access in layers: network reachability, then cryptographic per-connection identity, then verifiable credentials. A model that can persist doesn't get a door to persist against.