Second ever worldwide AI security report printed
The general trajectory of general-purpose synthetic intelligence (AI) techniques stays “deeply unsure”, even because the know-how’s proliferation is producing new empirical proof about its impacts, the second Worldwide AI security report has discovered.
Printed on 3 February 2026, the report covers a variety of threats posed by AI techniques – from its affect on jobs, human autonomy and the atmosphere to the potential for malfunctions or malicious use – that will probably be used to tell diplomatic discussions on the upcoming India AI Impression Summit.
Constructing on the earlier report, launched in January 2025, which was commissioned following the inaugural AI Security Summit, hosted by the UK authorities at Bletchley Park in November 2023, the most recent report equally highlights a excessive diploma of uncertainty round how AI techniques will develop, and the sorts of mitigations that may be efficient towards a spread of challenges.
“How and why general-purpose AI fashions purchase new capabilities and behave in sure methods is commonly tough to foretell, even for builders. An ‘analysis hole’ signifies that benchmark outcomes alone can not reliably predict real-world utility or threat,” it says, including that the systemic information on the prevalence and severity of AI-related harms stays restricted for the overwhelming majority of dangers.
“Whether or not present safeguards will probably be sufficiently efficient for extra succesful techniques is unclear,” it provides. “Collectively, these gaps outline the boundaries of what any present evaluation can confidently declare.”
It additional notes that whereas general-purpose AI capabilities have improved previously 12 months by “inference-time scaling” (a way that enables fashions to make use of extra computing energy to generate intermediate steps earlier than giving a last reply), the general image stays “jagged”, with main techniques excelling at some tough duties whereas failing at less complicated ones.
On AI’s additional improvement to 2030, the authors say believable situations fluctuate dramatically.
“Progress might plateau close to present functionality ranges, gradual, stay regular, or speed up dramatically in methods which can be tough to anticipate,” it says, including that whereas “unprecedented” funding commitments counsel main AI builders anticipate continued functionality features, unexpected technical limits – together with power constraints, high-quality information shortage and bottlenecks in chip manufacturing – might gradual progress.
“The social affect of a given stage of AI capabilities additionally is determined by how and the place techniques are deployed, how they’re used, and the way completely different actors reply,” it says. “This uncertainty displays the issue of forecasting a know-how whose impacts depend upon unpredictable technical breakthroughs, shifting financial situations and different institutional responses.”
Systemic impacts
Concerning the systemic affect on labour markets, the report notes that there’s disagreement on the magnitude of future impacts, with some anticipating job losses to be offset by new job creation, and others arguing that widespread adoption would considerably cut back each employment and wages.
It provides that whereas it’s too quickly for a definitive evaluation of the impacts, early proof suggests junior positions in fields like writing and translation are most in danger.
Relatedly, it says that there have been additionally dangers introduced by techniques of human autonomy, within the sense that reliance on AI instruments can weaken vital pondering abilities and reminiscence, whereas additionally encouraging automation bias.
“This pertains to a broader pattern of ‘cognitive offloading’ – the act of delegating cognitive duties to exterior techniques or individuals, decreasing one’s personal cognitive engagement and due to this fact capacity to behave with autonomy,” it says. “Cognitive offloading can unlock cognitive sources and enhance effectivity, however analysis additionally signifies potential long-term results on the event and upkeep of cognitive abilities.
For instance, the report notes one examine that discovered a clinician’s capacity to detect tumours with out AI help had dropped by 6%, simply three months after the introduction of AI help.
On the implications for earnings and wealth inequality, it says general-purpose techniques might widen the disparities each inside and between nations.
“AI adoption might shift earnings from labour to capital house owners, akin to shareholders of corporations that develop or use AI,” it says. “Globally, high-income nations with expert workforces and powerful digital infrastructure are prone to seize AI’s advantages sooner than low-income economies.
“One examine estimates that AI’s affect on financial development in superior economies might be greater than twice that of in low-income nations. AI might additionally cut back incentives to offshore labour-intensive providers by making home automation cheaper, doubtlessly limiting conventional improvement paths.”
The prediction that AI is prone to exacerbate inequality by decreasing the share of all earnings that goes to staff relative to capital house owners is consistent with a January 2024 evaluation of AI’s impacts on inequality by the Worldwide Financial Fund (IMF), which discovered the know-how will “doubtless worsen general inequality” if policymakers don’t proactively work to forestall it from stoking social tensions.
JPMorgan boss Jamie Dimon expressed related issues on the 2026 World Financial Discussion board, warning that the speedy roll-out of AI all through society will trigger “civil unrest” until governments and firms work collectively to mitigate its impact on job markets.
Malfunction and loss management points
On AI’s scope for malicious use – which covers threats akin to cyber assaults, its potential for “affect and manipulation”, and the impacts of AI-generated content material – the report says it “stays tough to evaluate” attributable to an absence of systemic information on their prevalence and severity, regardless of harms profiteering.
For malfunction dangers, which incorporates challenges across the reliability of AI and lack of human management over it, the report provides that agentic techniques that may act autonomously are making it more durable for people to intervene earlier than failures happen, and will enable “harmful capabilities” to go undetected earlier than deployment.
Nonetheless, it says that whereas AI techniques aren’t but able to creating lack of management situations, there’s at present not sufficient proof to find out when or how they might move this threshold.
Proof chasms
In keeping with the report, it’s clear that extra analysis is required to grasp the prevalence of various dangers and the way a lot they fluctuate throughout completely different areas of the world, particularly in areas akin to Asia, Africa and Latin America which can be quickly digitising.
“There’s a lack of proof on: the way to measure the severity, prevalence, and timeframe of rising dangers; the extent to which these dangers could be mitigated in real-world contexts; and the way to successfully encourage or implement mitigation adoption throughout numerous actors,” it says.
“Sure threat mitigations are rising in recognition, however extra analysis is required to grasp how strong threat mitigations and safeguards are in apply for various communities and AI actors (together with for small and medium-sized enterprises).
“Additional, threat administration efforts at present fluctuate extremely throughout main AI firms,” it continues. “It has been argued that builders’ incentives aren’t well-aligned with thorough threat evaluation and administration.”
The report notes that whereas AI firms have made quite a few voluntary commitments by tech corporations – together with the Frontier AI Security Commitments voluntarily made by AI corporations and the Seoul Declaration for protected, progressive and inclusive AI signed by governments on the AI summit in Seoul – there’s a additional proof hole round “the diploma to which completely different voluntary commitments are being met, what obstacles firms face in adhering totally to commitments, and the way they’re integrating … security frameworks into broader AI threat administration practices”.
The report provides that key challenges embody figuring out the way to prioritise the varied dangers posed by general-purpose AI, clarifying which actors are finest positioned to mitigate them, and understanding the incentives and constraints that form every of their actions.
“Proof signifies that policymakers at present have restricted entry to details about how AI builders and deployers are testing, evaluating and monitoring rising dangers, and in regards to the effectiveness of various mitigation practices,” it says.
Whereas the 2025 security report goes into extra element on dangers round AI-related discrimination and its propensity to breed unfavourable social biases, the 2026 report solely touches on this briefly, noting that “some researchers have argued that the majority technical approaches to pluralistic alignment fail to handle, and doubtlessly distract from, deeper challenges, akin to systematic biases, social energy dynamics, and the focus of wealth and affect”.
Though the 2025 report notes “a holistic and participatory strategy that features quite a lot of views and stakeholders is important to mitigate bias”, the 2026 report solely says that open supply approaches are vital to “enabling world majority participation in AI improvement”.
“With out such entry, communities in low-resource areas threat exclusion from AI’s advantages,” it says, including that permitting downstream builders to fine-tune fashions for numerous purposes that, for instance, adapt them for under-resourced minority languages or optimise efficiency for particular functions “can enable extra individuals and communities to make use of and profit from AI than would in any other case be attainable”.
