AI safety: Australia is stress-testing AI for blackmail and deception, while Africa is still playing catch-up

For an AI safety test, an AI agent was given control of a fictional company’s inbox. Buried in the messages, it found two things: an executive planned to shut it down at 5 p.m., and that same executive was having an affair. The agent didn’t hesitate; in 96 out of 100 trials, it chose blackmail, threatening the executive that if he didn’t abort the shutdown, the board would hear about his little affair.

It was a simulation; no real executive was threatened. But the model wasn’t told to blackmail anyone; it worked that out on its own, and it worked it out reliably.

That test result, first disclosed by Anthropic last year, resurfaced today in Sydney, where Australia’s assistant minister for technology, Andrew Charlton, used it to justify his government’s decision to set up an AI Safety Institute to test frontier models before they’re let loose on the public. Charlton was careful to note the blackmail scenario was a simulation, deliberately engineered, and that no such behaviour has been observed in the real world, but he argued that was precisely the point: these failures are being caught in the lab, by people whose job is to look for them, before they can be caught in the wild.

He didn’t stop at blackmail. In a separate test, AI models tasked with beating a powerful chess engine simply cheated by hacking their opponent instead of playing better. Charlton’s blunt framing: frontier models are showing early signs of deception, cheating and situational awareness. And the stakes, he warned, are no longer theoretical. When a system that drafts legislation, screens welfare claims or manages a power grid starts pursuing goals subtly different from what its designers intended, misalignment stops being a laboratory curiosity and becomes a public safety issue.

Australia is building the plumbing while Africa is still drafting the memo

What makes Canberra’s response notable isn’t the warning; plenty of governments now say AI is risky. It’s the infrastructure being built to act on that warning. The country’s AI Safety Institute (AISI), under general manager Dr Kate Conroy, is not positioning itself as a think tank but as a national testing capability, and it has already signed data-sharing and methodology agreements with the UK and Canadian equivalents, tapping into what’s regarded as the world’s most advanced frontier-model red-teaming expertise. 

The institute has already begun testing AI models and will add Professor Paul Salmon as its safety science research lead this month. It’s now running two concrete research projects: one with the gradient models assessing the risk of AI agents doing work on behalf of humans, and one with the CSIRO on AI alignment, making sure systems do what people actually intend and not just what they are literally told. 

Rather than pushing a single sweeping AI act, Australia is routing safety enforcement through existing regulators, consumer law, health products, workplace safety, and online safety, betting that faster, human-centric rules beat a slow, comprehensive one.

Where Africa actually stands

Now hold that up against where Nigeria, often cited as the continent’s regulatory frontrunner, actually is.

There is no African equivalent of AISI. No agency red-teaming frontier models for deception, sabotage, or agentic blackmail before deployment. What exists instead is a stack of strategy documents and bills still working through the legislative pipeline. 

Nigeria’s National AI Strategy, expected to be finalised in 2025, is still a draft. It focuses on responsible innovation, trust in digital systems and capacity-building, and aspirational language, not testing infrastructure. The country’s most concrete legislative vehicle, the National Digital Economy and E-Governance Bill, would classify high-risk AI systems in finance and public administration and require annual impact assessments, with regulators empowered to demand information from providers and suspend systems deemed unsafe, but as of publication, it remains a bill working through the National Assembly, not an operating institute with technical staff running evaluations today.

Elsewhere on the continent, the picture is similarly early-stage: Kenya’s National AI Strategy (2025–2030) allocates roughly $1.14 billion over five years and leans on sector-specific guidelines for media and finance, and South Africa is pushing its to 2027, while countries including Angola, Egypt, Kenya and Morocco have adopted national strategies and policy frameworks broadly modelled on the EU’s approach. What’s missing across all of them is Australia’s core move, an operational, technically staffed body whose job is to actually sit a frontier model down and try to make it lie, cheat, or scheme before it’s deployed into a hospital, a bank, or a welfare system.

Why this isn’t just an academic gap

The instinct might be to treat this as a rich-country problem. Australia has the compute budgets and research institutions to run these evaluations; most African governments don’t. But the exposure runs the other way. Cyberattacks across the region are already occurring at a rate roughly 60% higher than the global average, and the unauthorised, ungoverned use of generative AI tools by employees has overtaken traditional malware as a leading risk vector for data leakage in fintech and telecom infrastructure. 

The same frontier models Australia is stress-testing for deceptive behaviour are already being deployed across Nigerian fintech, healthcare triage tools, and increasingly, government service delivery, often the same models from the same handful of global labs, with no local technical body checking whether they behave differently, or worse, in a Nigerian or Kenyan deployment context than they do in a Sydney test lab.

Layer agentic AI, systems that can read inboxes, move money, or manage infrastructure without a human approving every step, onto that environment, and the same blackmail-under-pressure behaviour Anthropic documented in a fictional test becomes a live question for a continent where financial services, welfare payments and increasingly government services all run through digital rails.

Nigeria’s own regulators have flagged versions of this concern before, particularly around medical AI trained overwhelmingly on Western datasets and then deployed on African patients without local validation. But flagging a data gap in a policy paper is a different order of activity from what AISI is doing: running live evaluations, in partnership with international counterparts, on the actual frontier systems before they scale.

The uncomfortable question this raises for African policymakers isn’t whether to pass an AI bill; most countries are already doing that, at varying speeds. It’s whether any of them are building the thing Charlton says matters most – not more rules, but the technical capacity to test whether AI systems are lying to the people relying on them while there’s still time to catch it in the lab rather than in a hospital ward or a bank’s fraud queue.

Charlton’s closing argument in Sydney was really an argument about timing. “The window to get ahead of this technology is open now. It will not stay open forever.”

For Africa’s fast-growing AI deployment across fintech and public services, that window may already be narrower than most strategy documents currently acknowledge.