A shared photo on first glance may appear trivial. Provided platforms are transparent about how data is used, users can share according to individual thresholds for how much private life is public. However, at the macro level, when millions of individual shares are collated, who looks for patterns, and what can they do with that information in aggregate?
Consider a single listing: the photo type could be a copy or an original, a story or a post on any number of platforms: a twenty-something-year-old takes a selfie in a full-length mirror in a bedroom. Visible on the wall behind them are photos of friends, family, a dog and travel alongside a university graduation certificate. On the dressing table to the right, a series of personal items tell more of their story – a bottle of fragrance, a moisturiser, medication, a glass drinking bottle and a document with a government insignia.
On the surface, the photo looks innocuous, nothing more than common, normalised overexposure for the 21st Century. Let’s say this scenario is playing out on Depop, a global public social media platform full of fashion-conscious buyers and sellers. Manual doxxing is a breeze, but automated doxxing? Already done. The platform hosts millions of similar listings with geolocations, names and PII openly available. Depop’s own privacy policy confirms it: usernames, location, followers, listings – photographs, videos, prices, descriptions – are all public by design. The reality of 2026 is that once the OSINT door opens, it’s tricky to close.
AI Collapses the Cost of Collection
VLMs (Visual Language Models) have substantially shifted the OSINT (Open Source Intelligence) landscape. A 2025 evaluation of multimodal large language models found that the strongest models could localise street-level imagery with up to 49% accuracy within a 1-kilometre radius.
For investigators, this is excellent news. The capability means PII and pattern-of-life extraction from casual images can increasingly be automated at scale.
It comes with a cost asymmetry, though. Defenders are often confronting millions of individual behaviours, while collectors can potentially apply a singular pipeline for data exfiltratration.
Google has observed state-backed actors using AI for reconnaissance, target research, phishing preparation and data processing, including attempts to use AI to link personal attributes and derive information about individuals from structured datasets.
The implication is clear, as with many large model innovations, adversaries, including state-level actors, can weaponise intelligence collection at scale.
Collection Points
The Australian Signals Directorate and international partners stated that since 2022, following Russia’s invasion of Ukraine, Unit 26165 expanded targeting of logistics entities and technology companies involved in aid delivery and targeted internet-connected cameras at Ukrainian border crossings to monitor and track aid shipments.
GRU Unit 26165 (APT28) mass-compromised private and municipal IP cameras across Ukraine and NATO border states to monitor Western aid shipments.
In Iran, publicly available information can form part of the target persona. Mandiant documents APT42’s use of highly targeted social engineering, collection against individuals, and compromise of accounts belonging to employers, colleagues and relatives of initial victims. The oversharing century is fodder for threat actors like APT42/Charming Kitten, who mine targets’ social media and networks to construct personas for targeted spear-phishing campaigns and extend the collection to family members, widening the net of capture and the impact.
In China, the model can operate at an industrial scale. In March 2025, US indictments allege that i-Soon operated as part of a PRC hacker-for-hire ecosystem, working with at least 43 MSS/MPS bureaus across 31 provinces and municipalities.
Lastly, North Korean IT workers used stolen identities, false websites, social-media accounts, proxy computers and U.S.-based intermediaries to obtain legitimate remote employment. Once inside, workers gained access to company systems and, in some cases, stole sensitive information. The DPRK and the associated IT-worker tradecraft demonstrate how operations can exploit systems and relationships that defenders assume are benign.
The data harvesting pattern isn’t hypothetical for Western militaries. In late August 2026, the U.S Navy issued a written notice – ‘Epic Vigilance: Immediate actions for Force Protection and Personal Security’ – warning that adversaries were executing a multi-domain campaign against the Department of the Navy’s (DON) personnel. The DON amounts to just shy of 350,000 active-duty personnel, 58,000 reservists and 210,000 civilian workers. As part of the notice, the DON urged personnel to ‘remove any information that could identify your connection to the DON.’ Any casual LinkedIn user will recognise the surprising frequency of broadcasting clearance and photos on the platform.
This method of intelligence collection via social media has been a noted risk for well over a decade now. The Strava heatmap release inadvertently revealed secret U.S and allied operational bases in conflict zones like Afghanistan, Iraq, Syria and Niger.
Despite the known risk, military personnel have continued to publicly share data on social media. The location of a French aircraft carrier was leaked following a sailor publicly recording a run on deck. Over 1,300 U.S Service Members in the Middle East shared their workouts – even with the respective departments of defence engaging in opsec training, banning the use of fitness tracking devices in designated areas and issuing strict geolocation controls.
Almost Owned
Part of the ongoing challenge is that privacy law often frames this as individual consumer harm (consent notice). Platform consent does not necessarily mean that every subsequent use of information is simply a matter of individual risk acceptance. For example, Depops privacy policy distinguishes between publicly visible user information, information collected by the platform, and different legal bases for processing but also notes granted permissions allow for use of content externally for promotional purposes.
The question is who, if anyone, owns the aggregate exposure. Consider this. Even where an organisational owner does exist, the model strains. The DON treats personnel exposure as a force-protection risk, replete with opsec training and device bans, and Strava shows those controls being undermined by individual behaviour. Even the best-resourced organisation on earth cannot fully eliminate exposure arising from individual behaviour.
Where to From Here?
There is no quick fix. Compromise on secure-by-design in favour of fast deployment and a tightening competitive market, combined with blunted education on cyber hygiene and online safety in a world anchored in social media, has left devices and their humans vulnerable.
This has been heightened by the rapid adoption of IoT devices, and compounded by AI. Sure, stripping the EXIF, blur-by-default and scraping resistance as a security control would be excellent steps forward by platforms, but relying on platform self-regulation is unlikely to be sufficient.
Aggregate civilian exposure belongs in national cyber risk assessments. Predictably, objection may take the form of an argument that this is privacy regulation’s job. However, the boundaries above show why. Privacy law primarily governs the relationship between a user and a platform, not a threat actor who can potentially aggregate millions of individual data points.
The Forcing Function
The aggregate of publicly available information about homes, interests and movement can reveal significant intelligence to an adversary – largely freely available – even when individual disclosures appear innocuous.
The APT28 campaign demonstrates the underlying principle at operational scale. Individually ordinary internet-connected cameras became useful because their feeds could be combined.
If we, as security professionals, keep treating civilian exposure as out of scope while mining it for OSINT investigations, we risk leaving the aggregate intelligence value of that exposure primarily to the adversary. The alternative is to put it on the record and in our threat models
Leave a Reply