Skip to content
Briefings are running a touch slower this week while we rebuild the foundations.See roadmap
AI: Jobs, Power & Money
28MAR

GPT-5.5 clears 32-step attack chain; two models in five days

4 min read
19:20UTC

The UK AI Safety Institute confirmed on 6 May 2026 that GPT-5.5 cleared the 32-step autonomous cyber attack chain benchmark, becoming the second model to do so after Claude Mythos, with AISI's Frontier AI Trends Report recording frontier cyber capability doubling every four months.

← Back to The AI jobs data contradicts itselfJump to analysis ↓
EconomicDeveloping
Key takeaway

GPT-5.5 and Claude Mythos both cleared the 32-step attack chain within five days, making it a class-level capability.

The AISI (UK AI Safety Institute, the UK Government body established in November 2023 to evaluate frontier AI capabilities) published its Frontier AI Trends Report on 6 May 2026, confirming that GPT-5.5, OpenAI's frontier language model, cleared the 32-step autonomous cyber attack chain benchmark known as "The Last Ones" (TLO). GPT-5.5 achieved 71.4% on the expert cyber suite and solved the TLO benchmark end-to-end in 2 of 10 attempts. GPT-5.5 is the second model to clear the benchmark; Anthropic's Claude Mythos cleared the same threshold on 1 May 2026 .

AISI's report assesses frontier cyber capability as doubling every four months. The 32-step attack chain benchmark triggered an emergency convening of five Wall Street bank CEOs by Treasury Secretary Bessent and Fed Chair Powell in April 2026, after Claude Mythos became the first model to clear it . Five days after Mythos cleared the threshold on 1 May, GPT-5.5 repeated the result on 6 May.

Two separate frontier models clearing the same benchmark within five days changes the nature of the risk assessment. A single-model capability that is deliberately kept under restricted access (as Mythos Preview was, through Project Glasswing with access limited to twelve partners) can be managed through deployment controls. A benchmark cleared by two models from different organisations within five days of each other is no longer a single-deployment governance question; it is a class of capability that has arrived across the frontier.

AISI's four-month doubling rate, if it holds, means the next iteration of the same benchmark class will be cleared by a broader set of models within months. AISI's evaluation function, confirming which models have crossed which thresholds, is doing the work that no other institution with formal standing has yet stepped in to perform. Whether that evaluation function informs regulatory action in any jurisdiction within a timeframe that is operationally relevant to the doubling rate is the question AISI's report leaves open.

Deep Analysis

In plain English

On 6 May 2026, the UK AI Safety Institute published a report confirming that GPT-5.5, an AI model from OpenAI, had passed a difficult cyber security test. The test involves completing 32 consecutive steps in a simulated computer attack without human help. Only one other model, Anthropic's Claude Mythos, had passed this test before, and it did so just five days earlier. The report also said that frontier AI models are getting twice as capable at cyber attacks every four months. This matters because cybersecurity has been one of the sectors least affected by AI job cuts, because the technology was not yet reliable enough to replace human security analysts. A confirmed four-month capability doubling, if it continues, changes that assumption within 12-18 months.

Deep Analysis
Root Causes

The relevance to the AI-jobs topic is the labour market implication of the four-month capability doubling that the body does not name explicitly. Cybersecurity is one of the few technology employment sectors that has remained at full employment or near it throughout the 2025-2026 restructuring cycle: demand for human cyber analysts has been resilient because autonomous cyber capability has been below the operational reliability threshold.

The AISI confirmation that two models have now cleared the 32-step benchmark changes the supply side of the cybersecurity labour market: if frontier AI models can autonomously execute 32-step attack chains at a 20% success rate today, security operations centres will face pressure to reduce analyst headcount as that rate improves toward operational reliability.

Lloyd's of London cyber underwriters base pricing partly on the probability of a successful autonomous attack on insured infrastructure, making the AISI benchmark confirmation an underwriting input. A confirmed four-month doubling of frontier cyber capability is an underwriting input that re-prices coverage upward, affecting every enterprise that carries cyber insurance.

What could happen next?
  • Risk

    Security operations centre headcount justifications at major banks and critical infrastructure operators begin to face the same AI displacement pressure as software engineering if the four-month doubling sustains through 2026.

    Medium term · 0.6
  • Consequence

    Cyber insurance pricing must incorporate the AISI-confirmed capability doubling as a loss probability input; Lloyd's and specialist cyber underwriters will reprice 2027 coverage upward.

    Short term · 0.7
  • Risk

    The absence of a second interagency convening following GPT-5.5's clearance, five days after the first model cleared the same benchmark, suggests the institutional alarm response is not scaling with the capability curve.

    Immediate · 0.65
Sources:Bloomberg / CFO Dive·Bloomberg / CFO Dive
Mentions:GPT-5.5 →Anthropic →OpenAI →Federal Reserve →Claude Mythos Preview →Project Glasswing →Codex 5.3 →Claude Opus 4.6 →RAND Corporation →UK AI Security Institute →
First Reported In

Update #9 · GitLab signs the manifesto, Brussels backs out

Bloomberg / CFO Dive· 15 May 2026
Read original →
Causes and effects
Caused by
UK vacancies frozen at 721,000 for six months
711,000 April reading breaks downward from six-publication plateau that had been the covered story
Occurred 10 Apr 2026
Read story →
This Event
GPT-5.5 clears 32-step attack chain; two models in five days
Two separate frontier models clearing the same 32-step attack chain within five days of each other confirms the benchmark is no longer a single-model capability threshold, removing the assumption that frontier cyber risk remains isolated to a single controlled deployment.
Different Perspectives
Entry-level and displaced workers globally
Entry-level and displaced workers globally
Challenger's 69% April hiring-plan collapse means the entry-level market contracted faster than announced layoff figures indicate. Workers aged 22-25 in AI-exposed occupations show a 16% employment decline since late 2022; the Stanford JOLTS analysis puts the real AI labour impact at 34 times the declared Challenger count.
Chinese courts and regulators
Chinese courts and regulators
The Hangzhou Intermediate People's Court upheld in April that employers cannot dismiss for AI cost reasons without offering retraining, confirming the Beijing court's December 2025 precedent under Labour Contract Law Article 40. Chinese workers now hold the only binding, judicially tested AI employment protections in any major jurisdiction.
Investors
Investors
Markets are rewarding the AI restructuring trade. Cloudflare reported record revenue alongside its 20% cut; the companies endorsing S.3339, a commission study bill with no enforcement mechanisms, are the same companies executing the restructurings the commission would study.
EU member states and Council
EU member states and Council
The Council's non-binding encouragement clause won the 7 May Digital Omnibus trilogue, dropping 18 months of work toward a binding employer AI literacy obligation. The outcome reflects the trade-off member states made: regulatory flexibility for employers over enforceable worker protections.
AI-era tech CEOs
AI-era tech CEOs
Cloudflare's Matthew Prince framed the 1,100-job cut as 'defining how a high-growth company operates in the agentic AI era', not a cost reduction. GitLab's Bill Staples published the most candid CEO-signed thesis of the cycle: agents will plan, code, review, deploy, and repair.
US tech workers and organised labour
US tech workers and organised labour
SAG-AFTRA's failure to win the Tilly tax, following WGA's settlement without AI training payment, confirms that organised creative workers cannot secure royalty mechanisms for AI-generated characters. For software workers, GitLab's 60-team structure eliminates the managerial co-ordination layer without replacing it with equivalent roles.