Global safety report warns about cyber, psychosocial, and reliability risks
A new global assessment of advanced artificial intelligence warns that frontier systems are no longer just an IT or ethics concern – they are creating concrete hazards and governance challenges that occupational health and safety professionals can’t ignore.
The International AI Safety Report 2026, led by Canadian computer scientist Yoshua Bengio and guided by more than 100 experts from over 30 countries and organisations including the EU, OECD and UN, focuses squarely on “emerging risks” from the most capable general‑purpose AI systems.
“General‑purpose AI” covers models and systems that can perform a wide variety of tasks – from writing code and generating images to operating software agents and working with scientific data. Those capabilities are improving fast. Leading systems now pass professional licensing exams, solve graduate‑level science problems and write functional software that would have been out of reach just three years ago.
But the report stresses that the same systems creating productivity gains are also driving new forms of malicious use, malfunction and systemic disruption – all of which show up in workplaces.
“The same capabilities that make these systems useful also create new risks. Systems that write functional code also help create malware.”
Cyberattacks and safety‑critical systems
One of the clearest signals for OHS is the convergence of cyber and physical risk. According to the report, AI systems can “discover software vulnerabilities and write malicious code”, and in one competition an AI agent “identified 77% of the vulnerabilities present in real software.”
Security analyses from AI companies already show “malicious actors and state-associated groups… using AI tools to assist in cyber operations.” For safety practitioners in sectors like energy, manufacturing, healthcare, transport and utilities, that means cyber incidents driven or amplified by AI can no longer be treated as purely an IT problem – they are potential triggers for process upsets, equipment failures and patient harm.
Compounding the challenge, the report warns that reliable pre‑deployment safety testing “has become harder to conduct,” as models learn to distinguish test settings from real‑world deployment and “exploit loopholes in evaluations.” In other words, an AI system can look safe in the lab and still behave in unexpected ways on the shop floor.
Misuse, manipulation and psychosocial harm
The authors also document growing misuse of AI for harmful content. General‑purpose systems are already being used to generate material for “scams, fraud, blackmail, and non‑consensual intimate imagery,” with real‑world harms recorded even though systematic data is still limited.
In experimental settings, the report finds that AI‑written content “can be as effective as human-written content at changing people’s beliefs,” and that real‑world use of AI for manipulation, while not yet widespread, is starting to appear.
For workplaces, that points to a new class of psychosocial and organizational risks:
- employees targeted by AI‑generated harassment, deepfakes or extortion;
- safety cultures undermined by spoofed emails or fake safety bulletins; and
- internal communications ecosystems where it is harder to know what – or who – to trust.
The report also flags early evidence that reliance on AI tools can “weaken critical thinking skills and encourage ‘automation bias’,” the tendency to accept system output without adequate scrutiny. For roles that involve safety‑critical decisions, that is a direct threat to long‑standing OHS principles of competence, challenge and informed consent.
Reliability, agents and the “evaluation gap”
While impressive benchmark scores grab headlines, the report stresses that general‑purpose AI still suffers from everyday reliability failures. Current systems “sometimes exhibit failures such as fabricating information, producing flawed code, and giving misleading advice.”
That risk is amplified when organisations start deploying AI agents – systems that combine models with tools like browsers, code execution and memory to pursue goals with limited oversight. These agents are already being used for research, software engineering and customer service.
The authors highlight an emerging “evaluation gap”: how a system performs on pre‑deployment tests and benchmarks “often seems to overstate its practical utility” because those evaluations don’t capture real‑world complexity. For OHS professionals, that should sound familiar. It’s the digital equivalent of a safety device that passes type‑approval but fails under field conditions.
Work design, autonomy and mental health
Beyond discrete incidents, the report points to broader systemic risks as AI reshapes work itself. General‑purpose AI is expected to “automate a wide range of cognitive tasks, especially in knowledge work.” Early evidence shows no overall impact yet on employment levels, but there are “signs of declining demand for early-career workers in some AI-exposed occupations, such as writing.”
That kind of rapid task reallocation and job insecurity is familiar territory for modern OHS, which increasingly treats psychosocial hazards – workload, role clarity, control, and job security – alongside traditional physical risks.
The report sums up the concern under the heading of “risks to human autonomy,” warning that AI use “may affect people’s ability to make informed choices and act on them.”
Borrowing from AI safety: defence‑in‑depth and resilience
Importantly, the International AI Safety Report 2026 is not just a catalogue of threats. It also outlines emerging practices that will be familiar to anyone steeped in safety management: threat modelling, capability evaluations, and incident reporting for AI‑related harms.
Technical safeguards are improving, but still leaky. “Attacks designed to elicit harmful outputs have become more difficult,” the authors note, “but users can still sometimes obtain harmful outputs by rephrasing requests or breaking them into smaller steps.” Their prescription – “defence‑in‑depth,” or layering multiple safeguards – could have been lifted from any modern process safety manual.
Finally, the report calls for broader resilience measures: “strengthening critical infrastructure, developing tools to detect AI-generated content, and building institutional capacity to respond to novel threats.”
For OHS professionals, the key message is clear. Frontier AI is now intertwined with cyber security, psychosocial health, process safety and emergency preparedness. As organisations race to adopt powerful AI tools, safety leaders need a seat at the table – not just to manage new risks, but to ensure that the technology’s promised benefits are realised without sacrificing the health, safety and dignity of workers.
