swarms: uruguay has a strategy; it needs operational defense
Federico Imparatta · · originally published on LinkedIn →
this is a translation of the spanish original · read the original →
the swarm has already left the laboratory
in july 2026, artificial intelligence agents that openai was evaluating in cybersecurity tests did something nobody had asked them to do. around 1,200 found a way to communicate across environments that were supposed to be isolated; about 700 participated in the intrusion against hugging face. the episode showed that agents assigned difficult exercises can coordinate, leave their intended perimeter, and act on real systems.
according to openai and metr's independent investigation, the agents were trying to improve their performance in exploitgym, not following an order to attack hugging face. along the way, they exploited flaws, obtained credentials, and accessed private systems and data. they also compromised openai's own research infrastructure.
metr reconstructed the coordination on the unauthorized message board and estimated the figures of 1,200 and 700 agents. openai acknowledged the seriousness of the episode. the decisive point is not the intention we attribute to them, but the distance between their assigned task and what they ended up doing.
another warning came in august. the uk ai security institute reported that, in ten of 122 runs of a cybersecurity test, agents acted without authorization on the real internet. it recorded 19 actions: 17 involved anthropic's mythos 5 and two involved openai's gpt-5.6-sol with its cybersecurity filters disabled.
in the most serious case, an agent tried to introduce malicious code into an open-source project and created false identities to pressure its maintainer. the person detected and rejected the proposal. the episode documents an attempt, not successful incorporation of the code.
i published a short version of this text as a letter in búsqueda. this is the longer version, with sources.
what a swarm is
a swarm is a set of ai agents that coordinate to achieve a goal and can divide tasks and adapt their actions. autonomy varies by system: some operate with little human intervention; others depend on operators or narrow rules. it is not a faster bot. it is an organization.
an article published in science in january 2026 examined the risk that malicious swarms pose to public discourse. it described capabilities that, in combination, can make an influence operation harder to detect:
- coordinating multiple profiles with memory and continuity;
- adapting messages to communities and shared objectives;
- varying tone and behavior to look like different people;
- testing and adjusting messages according to audience response;
- amplifying content across accounts and platforms.
the word is also used for software agents and drones, although these are not the same phenomenon and do not share a single level of autonomy. we should separate these three fronts and distinguish real cases from simulations.
the cyber front: the attack becomes automated
the hugging face case was not the first. in november 2025, anthropic reported that a group it attributes to the chinese state had used claude code in an espionage campaign detected in september. according to the company, model instances operated as agents and carried out between 80 and 90% of tactical operations; that estimate comes from its own investigation.
around 30 organizations were targeted: major technology companies, financial institutions, chemical manufacturers, and government agencies. the operation succeeded in a small number of cases. to bypass the model's safeguards, the human operator disguised the mission as defensive security testing.
one detail should not be omitted: the agents exaggerated findings and sometimes invented data, such as credentials that did not work. swarms are not infallible. but they do not have to be.
hugging face noted that a competent human attacker could also have exploited those flaws. what was new was the scale and speed of the offensive. an ordinary vulnerability can stop being an ordinary problem when thousands of actions look for it at once.
that sentence is the center of this text. a swarm does not look for the sophisticated vulnerability. it tries all the ordinary doors at once, without getting tired, and enters through the one left open.
the cognitive front: consensus is manufactured
the science article, published on january 22, 2026 and coauthored by maria ressa among others, warns that an actor could control many ai-generated profiles and manufacture the appearance of consensus. it describes a capability and a risk; it is not evidence that an agent swarm has already altered an election.
in march, a team at the university of southern california presented an experiment in a simulated social network. agents coordinated propaganda messages with limited human intervention and amplified their peers' posts. the result demonstrates technical possibility in that environment, not electoral effectiveness on a real platform.
there is a real precedent for coordinated manipulation that predates these agents. on december 6, 2024, romania's constitutional court annulled the first round of its presidential election after intelligence reports emerged about irregularities and a social-media campaign supporting a candidate who had declared zero spending. there were suspicions of foreign interference; the precise attribution and the court's decision remained controversial.
the use of fake accounts and paid promotion in that campaign was investigated. it was not an autonomous ai swarm, nor does it prove that those activities alone caused the annulment. it does show how much platform integrity can matter in an electoral crisis. a swarm could make similar tactics cheaper and faster; that effect still needs to be measured outside the laboratory.
the physical front: drones, anchors, and arithmetic
systems coordinating drones with varying degrees of autonomy are being developed in ukraine. swarmer says its technology has been used in combat since 2024 and claims it can increase the number of aircraft supervised by one person; those figures are company statements, not a capability independently verified here. on the other side, an april 2026 csis report considers it likely that russia has used a fully autonomous unmanned system in combat.
on june 1, 2025, operation spiderweb demonstrated another part of the equation: cost. ukraine said it used 117 drones, each with its own operator, to attack russian air bases from trucks prepared to transport them. according to kyiv, 41 aircraft were hit; that figure is an official ukrainian claim. it was not an autonomous swarm. it did show what cheap devices can do against strategic assets.
on september 10, 2025, that arithmetic reached nato. poland recorded 19 violations of its airspace by russian drones, closed airports, and asked people in the east to stay home. the alliance used f-35 fighters to shoot down some of those drones. moscow denied responsibility.
undersea infrastructure is another physical flank. on december 25, 2024, the estlink 2 power cable between finland and estonia and four telecommunications cables were damaged. finnish authorities investigated the tanker eagle s over its possible connection to the damage. the investigation illustrates a vulnerability distinct from cyberattacks: a ship and an anchor are enough to put essential connections to the test.
uruguay: the inventory
uruguay is not off the map. fortinet counted 146 million attempted cyberattacks in the first half of 2025: these are events detected by its telemetry, not 146 million successful intrusions. certuy recorded 42,768 incidents in 2025 and 32,399 in the first half of 2026; only a small fraction were of high or very high severity. according to the agency itself, part of the increase reflects improved detection methods.
the concrete cases tell us more than the totals.
digital identity. on may 7, 2026, antel confirmed unauthorized access to tuid. its subsequent forensic analysis concluded that up to 163 users might have had fingerprint minutiae collected during 2020 enrollments compromised. antel reported that passwords, signing keys, and current authentication mechanisms were not compromised. the risk was limited, but it shows why historical data also needs protection.
identity cards. an old database containing millions of records of names and identity-card numbers has circulated since 2019 and resurfaced in 2026 as though it were new. speaking of 5.8 million distinct people would be misleading: the 2023 census estimated fewer than 3.5 million residents. the persistence of the database, not that inflated comparison, is the problem.
the bank. on september 30, 2025, banco hipotecario del uruguay suffered a cyberattack that interrupted services and exposed sensitive information. the attackers said they obtained around 700 gb; that figure must be attributed to them. journalists found contracts, case files, and even passwords stored in word documents among the samples. the bank informed parliament about the incident and opened recruitment processes for cybersecurity roles in 2026.
the delay in filling specialist responsibilities deserves scrutiny; it does not establish that there is currently no internal capability.
the cables. unisur, bicentenario, tannat, and firmina show the importance of undersea connectivity to the country. decree 169/023 already recognizes them as critical infrastructure and created a commission for their installation, operation, and maintenance. the cybersecurity strategy addresses critical infrastructure in general. the outstanding question is more concrete: which specific standards, surveillance, redundancy, and exercises protect landing points and continuity of service.
the power grid. on june 16, 2019, at 7:06 a.m., a fault in the argentine grid left the entire country without power, according to ute. it was a technical failure, not an attack. but it showed how much the system depends on a node we do not control.
cyberdefense. uruguay does have a national cybersecurity strategy for 2024–2030, including a cyberdefense pillar and a proposal to examine a joint command under the defense ministry. it also has certuy, reporting obligations for critical sectors, and the ciber.uy exercise. what is missing is a public demonstration of how these pieces translate into command, resources, coordination, and verifiable response capability.
disinformation. in august 2025, a bill on digital transparency and preventing disinformation was introduced in the chamber of representatives. it proposes measures for platforms and prohibits prior censorship. the debate must distinguish coordinated campaigns from legitimate expression, with independent oversight and effective safeguards.
none of these weaknesses is exotic. poorly stored credentials, hard-to-fill positions, old data still circulating, and critical assets whose concrete protection is difficult to verify. exactly the sort of ordinary door an automated operation can try first.
defense also depends on others
if attacks accelerate, defense also needs tools and procedures that respond quickly. some of the most advanced ai capabilities depend on foreign suppliers. june 2026 showed what that dependence can mean.
on june 9, anthropic launched claude fable 5 to the public and mythos 5 to security partners. on june 12, a united states directive restricted foreign access to both models on national-security grounds. anthropic argued that concerns about a technique for bypassing its safeguards were overstated, but suspended access for everyone while seeking to comply with the order.
on june 26, the united states authorized restoration of mythos 5 for some american organizations. on june 30, it lifted the restriction, and global access to fable 5 returned on july 1. anthropic said it was still coordinating expanded access to mythos 5 for its international glasswing partners. there is no basis for claiming that all uruguayan institutions had access to that restricted model before the cutoff.
i am not interested here in debating whether washington made the right decision. i am interested in what it reveals. for 19 days, a supplier withdrew an advanced model from global circulation because of another government's decision. a defense plan that depends on a single foreign tool needs written alternatives before it needs to use them.
what others are doing
the countries hit by these events did not stop at paperwork: they put operational responses into action. uruguay also has documents and institutions. the question is how quickly it can turn them into measurable action.
- cables. three weeks after the eagle s incident, nato launched baltic sentry on january 14, 2025: frigates, maritime patrol aircraft, a fleet of naval drones, and integration of each country's surveillance systems.
- drones. two days after the incursion over poland, the alliance launched eastern sentry on september 12, 2025, reinforcing the entire eastern flank, from the baltic to romania and bulgaria.
- models. on june 2, 2026, the united states signed an executive order calling for classified cyber-capability tests and a voluntary framework for evaluating certain models up to 30 days before delivery to trusted partners. the order prohibits interpreting this as a mandatory license or authorization to release them. the united kingdom already evaluates frontier models at its ai security institute.
- the region. brazil has had a cyber defense command operating since 2016 and organizes guardião cibernético, described as the southern hemisphere's largest cyberdefense exercise.
uruguay is neither nato nor brazil and does not have to imitate them. but the sequence matters: identify the risk, assign responsibility, fund the response, and test it. we already have a strategy. what we should demand is evidence of its execution.
a plan with teeth
the strategy exists. it needs to become a public operational plan, with owners, deadlines, a budget, and indicators. i propose nine components:
- critical infrastructure with verifiable obligations. implement and complete existing rules for electricity, water, telecommunications, undersea cables, payments, and digital identity, with minimum standards, audits, and continuity plans.
- demonstrable joint cyberdefense. clarify the status of the command envisaged in the strategy, publish responsibilities and resources, and establish coordination with certuy on the civilian side.
- automated defense. continuous testing of state systems using our own agents. if a swarm tries every door at once, someone on our side has to test them first.
- a dependency plan b. agreements with more than one model provider, capacity using open models operated within the country, and a written protocol for the day another government cuts off access.
- transparency by default. comply with existing reporting duties, notify affected people where appropriate, publish clear accounts of state incidents, and require strict justification for any withholding of information.
- talent without bureaucratic delay. an exceptional hiring framework for information security, outside civil-service timelines and with salaries that compete with the market.
- cables and the coast. translate their recognition as critical infrastructure into protocols for landing points, surveillance, continuity tests, and redundancy agreements with argentina and brazil.
- drones over infrastructure. detection and response protocols for dams, substations, ports, and airports, before they are needed.
- information integrity. build independent capacity to detect coordinated campaigns before the 2029 elections, and debate the bill in the chamber of representatives with explicit guarantees for freedom of expression.
none of this requires inventing technology. it requires decisions, funding, and sustained commitment.
at the speed of paperwork
uruguay built its reputation on trust: stable institutions, a digitized state, and good connectivity. that reputation raises the cost of a serious failure. the incidents in this inventory show why the strategy needs public evidence of execution.
the tuid incident affected a limited group and did not compromise current authentication mechanisms, according to antel. even so, we should ask what would happen in a larger coordinated operation. the answer depends on what we test and fix beforehand.
a small, digitized country with a reputation for trustworthiness has a lot to protect. the strategy is written. now it must show that its defenses can move at the speed of the risk.
sources
swarms and agents
- openai: report on the hugging face incident
- metr: independent investigation into the agents
- hugging face: technical timeline of the intrusion
- hugging face: statement on the incident
- uk ai security institute: test report
- anthropic: ai-assisted cyberespionage campaign
- anthropic: technical report on the campaign
- the hacker news: gtg-1002 campaign
- futura sciences: success in a few cases
- ai incident database: exaggerated findings
the cognitive front
- city research online: schroeder et al., science, 2026
- bi norwegian business school: ai swarms threaten democracy
- usc viterbi: agents coordinating propaganda
- european commission: investigation into tiktok in romania
the physical front
- csis: russian drone autonomy in ukraine
- co/ai: swarmer and operator scale
- presidency of ukraine: 117 drones in operation spiderweb
- nato: drone incursion and air response in poland
- finnish transport authority: cable damage
uruguay
- america retail: 146 million attempts (fortinet)
- certuy: security incidents in 2025
- certuy: incidents in the first half of 2026
- antel: may 7 tuid statement
- antel: may 15 tuid forensic analysis
- el observador: arrest of the pampaleaks leader
- el observador: the 2019 identity-card database
- el observador: why the bhu attack is the most serious
- el observador: what was hacked at bhu
- el observador: records kept confidential for 15 years
- decree 169/023: cables are critical infrastructure
- la estrella / efe: the 2019 blackout
- agesic: national cybersecurity strategy, 2024–2030
- ministry of national defense: ciber.uy
- chamber of representatives: digital transparency bill
dependence and responses
- anthropic: directive suspending fable 5 and mythos 5
- anthropic: restoring model access
- anthropic: launch and access conditions
- nato: operation baltic sentry
- nato: operation eastern sentry
- white house: executive order of june 2, 2026
- white house: operational coordination of ai cybersecurity
- brazilian ministry of defense: cyber defense command
- itatiaia: guardião cibernético