Four Rabbits

 · 13 min read
 · Nulla Verba
Table of contents

The last post found the clinical evidence base for aluminium adjuvant safety: 84 adults in two trials.1 But regulators don't rest their case on clinical trials alone. They point to pharmacokinetic modelling: mathematical models tracking how aluminium is absorbed, distributed, and eliminated after injection. The most cited model is from the US FDA. Its conclusion: the body burden of aluminium from vaccines poses "extremely low risk."2

What does that model rest on?

The Citation Chain

The UK's MHRA defers to "international evidence" on aluminium adjuvant safety, meaning the EMA and WHO.14 Italy's AIFA, as an EU national agency, implements EMA marketing authorisations directly. Both paths lead to the European Medicines Agency.

The EMA, questioned directly in 2019, replied that it has "no specific safety studies comparing aluminium alone with an inactive comparator." No new studies were needed, the EMA explained, because there were "no new safety issues reported after decades of use."3 For the pharmacokinetic case, it pointed to the WHO.

In 2012, the WHO's Global Advisory Committee on Vaccine Safety reviewed and endorsed what it called "the US FDA risk assessment model."4

That model is Mitkus et al. 2011. It enrolled no subjects. Its empirical inputs come from two earlier studies: Flarend et al. 1997 for how aluminium adjuvant is absorbed after injection, and Priest et al. 2004 for how aluminium is retained in the body once absorbed.

The chain:

MHRA / AIFA → EMA → WHO GACVS 2012 → FDA Mitkus 2011 → Flarend 1997 + Priest 2004

Each link looks solid from above. The strength depends on the bottom.

The Study at the Bottom

Flarend et al. 1997 is the only published study measuring the absorption of injected aluminium adjuvant in any species.5 It used radioactive ²⁶Al to track where adjuvant aluminium goes after intramuscular injection.

Six female New Zealand White rabbits. Two received aluminium hydroxide (AH). Two received aluminium phosphate (AP). One received intravenous aluminium citrate as reference. One served as contamination monitor.

The key comparison, how much adjuvant aluminium enters the blood, rests on two rabbits per adjuvant type, observed for 28 days:

Rabbit Adjuvant Blood absorption Urinary excretion
AH #1 Al hydroxide 13% 5.0%
AH #2 Al hydroxide 22% 6.2%
AP #1 Al phosphate 47% 10%
AP #2 Al phosphate 55% 33%

The two AP rabbits excreted 10% and 33% of the dose in urine, a 3.3-fold difference. With two animals per group, there is no way to know whether this spread is typical or extreme.

Then there are the missing samples.

Bone: Lost for all rabbits during chemical preparation. The authors wrote that "a significant fraction of the injected aluminum was probably incorporated into bone." Bone is a primary aluminium sink. The study most cited for tissue distribution data has no bone data.5

Brain: Lost for AP rabbit #2, the animal with the highest blood aluminium levels. The one brain most worth measuring is the one they lost. The three brains they did measure all contained aluminium at 28 days.6

Injection site: Not sampled. The depot theory (that aluminium sits at the injection site and slowly dissolves into the blood) centres on what happens there. Flarend did not measure it.5

Draining lymph nodes: Not sampled. The "lymph node" values in the paper come from mesenteric (intestinal) nodes, not the nodes draining the injection site.7

The authors wrote: "a much longer study would be needed to determine the final elimination of the aluminum from these adjuvants." They noted a human radiotracer study "should be possible since the amount of radioactivity is negligible."

Neither study was done. In 29 years, no one has reproduced the absorption kinetics that the model depends on.8

The Model

In 2011, FDA researchers published a mathematical model using Flarend's data, updating an earlier version from 2002; the underlying data did not change.132 The model takes 28-day rabbit absorption curves and extrapolates: constant absorption over 165 days for aluminium hydroxide and 56 days for aluminium phosphate. For retention, it uses data from one adult male volunteer tracked for twelve years after an intravenous dose of soluble aluminium citrate.9

The model compares the resulting body burden to a dietary safety limit: the ATSDR's (Agency for Toxic Substances and Disease Registry) oral Minimal Risk Level (MRL) of 1 mg/kg/day, derived from mice fed aluminium salts.11 Its conclusion: the body burden from the infant vaccine schedule stays below this threshold. "Extremely low risk."

The paper's admissions:

"the results of Flarend et al. are based on data from only two rabbits per adjuvant tested"

"it was primarily an exploratory investigation"

"Ideally, the results of that study should be confirmed using a larger number of animals, in order to increase our confidence in the results"

"the published retention function for aluminum is based on results for only one person"

"An infant monkey study could provide these data, however, given the present lack of evidence of harm due to the current aluminum levels, such studies may be a low priority"

Each limitation is acknowledged, then set aside. The paper says "ideally" and then explains why the ideal is unnecessary. The conclusion does not change.12 The implicit defence: the model's conclusion has held up across billions of doses without a clear safety signal. Whether that absence of signal constitutes evidence depends on what surveillance systems can detect.15

Priest, who conducted the single-volunteer study, wrote in his 2004 review: "Long-term retention of aluminium has only been studied in one volunteer."9 He measured six more adults, but only for five days. Even in that short window, retention varied from 16% to 37%, a 2.3-fold range.10 The long-term function comes from one adult male, measured with soluble aluminium citrate given intravenously. Mitkus applied it to particulate aluminium adjuvant injected intramuscularly in infants. Different form, different route, different species, different age.

The Point

Four rabbits. One person. One model.17

Component Source N Duration Species / Form
Absorption Flarend 1997 2 per adjuvant 28 days Rabbit, IM particulate
Retention Priest 1995/2004 1 12 years Human adult, IV soluble
Model Mitkus 2011 0 N/A Mathematical
Safety benchmark ATSDR 2008 N/A Oral, dietary Mouse

Two rabbits per group, with lost bone samples and one lost brain. The three brains they did measure all contained aluminium at 28 days. One adult male given a different form of aluminium by a different route. A model that enrolled no one. A dietary safety limit from mice, applied to injected particulate aluminium in infants.

A small exploratory study did what small exploratory studies do: opened a question. The question was never closed. It was modelled, endorsed, cited, and relabelled as settled. Each step was reasonable in isolation. Together, they produced a confidence the evidence does not support.

The pharmacokinetic case rests on four rabbits, one volunteer, and a model. The clinical case rests on 84 adults in two trials. Neither could have detected harm at any plausible rate, yet both are cited as evidence of safety.16 That leaves one claim: that decades of post-market surveillance would have caught a problem by now. Whether those systems were built to catch what these studies could not is worth examining.

If you spot an error in my reasoning, data, or sources, tell me. I'll correct it publicly.

  1. A 2022 systematic review found 102 trials that included an aluminium adjuvant arm. In 100 of those, the control arm also contained aluminium, vaccine antigens, or both. These designs test whether adding aluminium changes a vaccine's safety profile. Any effect of aluminium itself cancels out when both arms contain it. Only 2 of the 102 compared aluminium alone, with no vaccine antigens, against an inert placebo: Basavaraj 2014 (20 per arm) and Landrum 2017 (22 per arm). See 2 of 102 for the full breakdown. ↩

  2. Mitkus KJ, King DB, Hess MA, Forshee RA, Walderhaug MO. "Updated aluminum pharmacokinetics following infant exposures through diet and vaccination." Vaccine 2011;29(51):9538-43. doi:10.1016/j.vaccine.2011.09.124. All page references are to this paper. ↩↩

  3. Petersen SB, Gluud C. "Was amorphous aluminium hydroxyphosphate sulfate adequately evaluated before authorisation in Europe?" BMJ Evidence-Based Medicine 2021;26(6):285-289. doi:10.1136/bmjebm-2020-111419. PubMed 32763959. Based on direct queries to the EMA: ASK-50308 and ASK-53619. The EMA also stated that AAHS (the aluminium adjuvant in Gardasil-9 and other Merck vaccines) "was introduced without any prelicensure safety evaluation." ↩

  4. WHO Global Advisory Committee on Vaccine Safety. "Aluminium adjuvants." Weekly Epidemiological Record 2012;87(27):257-260. WHO WER. The committee also noted in 2004: "Validated animal models for adjuvant safety testing do not exist." ↩

  5. Flarend RE, Hem SL, White JL, et al. "In vivo absorption of aluminium-containing vaccine adjuvants using ²⁶Al." Vaccine 1997;15(12-13):1314-18. doi:10.1016/S0264-410X(97)00041-8. ↩↩↩

  6. Brain aluminium concentrations at 28 days: AH rabbits 0.009 and 0.008 μg/g; AP rabbit 0.019 μg/g (one value only; the other rabbit's brain was lost). Flarend 1997, Table 2. ↩

  7. Masson JD, Crépeaux G, Authier FJ, Exley C, Gherardi RK. "Critical analysis of reference studies on the toxicokinetics of aluminum-based adjuvants." J Inorg Biochem 2018;181:87-95. doi:10.1016/j.jinorgbio.2017.12.015. Masson also documents that the Mitkus model considers only dissolved aluminium ions and ignores particulate aluminium transported intact by macrophages, a mechanism unknown in 1997 but established by 2018. The paper argues this makes the oral-to-injected comparison fundamentally flawed: "the oral toxicity of soluble Al cannot be simply extrapolated to injected Al particles." Funded by ANSM (French drug agency), Île-de-France Region, and CMSRI. The CMSRI funding should be noted for transparency; the mathematical critique stands or falls on its own terms. ↩

  8. Multiple ²⁶Al tracer studies exist (Priest, Newton, Talbot) but all used oral or intravenous routes with soluble aluminium, not intramuscular adjuvant. The Hethey et al. 2021 curated ²⁶Al dataset explicitly excluded IM/SC data, noting the "complex and partly unknown absorption process." The most recent extension (Hartung, Wangorsch, Huisinga & Weisser 2025, Arch Toxicol 99:2379-2395, doi:10.1007/s00204-025-04031-1) added IM vaccination by estimating injection-site release rates from Flarend 1997 (rabbit), Weisser 2019/2020 (rat), McDougall 2016 (rat), and Verdier 2005 (monkey). Flarend remains in the dataset. The model assumes injection-site loss translates directly to absorption into blood (p.2386), which excludes macrophage-mediated transport to distant tissues without blood transit, the pathway described in The Aluminium Low-Dose Problem. Brain aluminium is modelled as passive diffusion from blood across the BBB only. When validated against Weisser 2019 rat data, the model predicted negligible brain Al, but all three vaccine groups showed statistically significant brain elevations. The authors dismissed these as "false positives" (p.2392), attributing the discrepancy to low control-group baselines. The model cannot predict the brain findings because it does not include the transport mechanism the brain findings imply. ↩

  9. Priest ND. "The biological behaviour and bioavailability of aluminium in man, with special reference to studies employing aluminium-26 as a tracer: review and study update." J Environ Monit 2004;6:375-403. doi:10.1039/b314329p. Direct quote: "To date, only one study has been undertaken to determine the long-term biokinetics of aluminium in the human body. This study was a single volunteer study." ↩↩

  10. The six additional volunteers were measured to 5 days only. Mean retention at 5 days: 25.7 ± 2.8%. Range: 16% to 37%. At 1 day post-injection, blood levels varied from 0.1% to 0.6% per litre (6-fold range). Priest noted that unpublished excretion data suggest the long-term retention half-time may be "in the region of 50 years," and that "the future trend in the retention is unpredictable." Priest 2004, Section 3.4. ↩

  11. Agency for Toxic Substances and Disease Registry. "Toxicological Profile for Aluminum." Atlanta, GA: US Department of Health and Human Services, 2008. Appendix A: ATSDR Minimal Risk Levels. The intermediate-duration oral MRL of 1 mg/kg/day is derived from a NOAEL of 26 mg Al/kg/day (Golub & Germann 2001, mice fed aluminium lactate), with uncertainty factor 100 and modifying factor 0.3 to account for aluminium lactate bioavailability (0.63%) being higher than typical dietary aluminium (0.1%). ↩

  12. "Lack of evidence of harm" is doing significant work here. It is the same framing examined earlier: absence of evidence functioning as evidence of absence. Mitkus uses "no evidence of harm" to justify not generating the evidence that could find it. ↩

  13. The Mitkus model updated an earlier one by Keith, Jones, and Chou (2002), all from the ATSDR. Keith assumed 100% immediate absorption of injected aluminium and used a higher MRL (2 mg/kg/day, from Golub 1989). Mitkus refined absorption to gradual dissolution (using Flarend) and adopted the 2008 MRL. The core empirical inputs, Flarend for absorption, Priest for retention, Greger for GI uptake, remained unchanged. Keith LS, Jones DE, Chou CH-SJ. "Aluminum toxicokinetics regarding infant diet and vaccinations." Vaccine 2002;20:S13-17. ↩

  14. MHRA. "MHRA reaffirms safety of childhood vaccination." GOV.UK. gov.uk. ↩

  15. This is the strongest defence of thin foundational evidence, and it deserves more than a sentence. Whether passive surveillance systems, with no unvaccinated comparator and 1-10% event capture rates, can detect the kinds of harm that 42 subjects cannot rule out is examined in a later post. ↩

  16. The largest study often cited here is Andersson et al. 2025, a Danish register-based cohort of 1.2 million children. The main analysis measures dose-response: hazard ratios per 1-mg increase in cumulative aluminium. This can detect whether more aluminium is worse than less. It cannot detect effects shared by all doses, for the same reason that comparing 20 cigarettes a day to 10 cannot detect harms common to all smokers. The cohort does include 15,237 children (1.2%) with no aluminium-adsorbed vaccines, but they anchor the low end of the continuous dose-response model rather than serving as a comparison group, and 82% had the lowest rate of GP visits, making direct comparison unreliable. ↩

  17. What these numbers mean at scale. The clinical evidence base (84 subjects in two trials) has 42 subjects per arm and no reported serious adverse events in the separated aluminium-vs-inert comparison. Under maximum ignorance (uniform prior, zero events in 42 subjects), the expected rate is approximately 1 in 44 and the 95% upper bound is roughly 1 in 14. The evidence is compatible with a serious harm rate anywhere from zero to 7%. To rule out a rate of 1 in 10,000, a randomised trial would need approximately 30,000 subjects per arm. The current evidence base is more than 700 times too small to answer that question. What the gap between 42 and 30,000 means for the population at risk is a question that deserves its own treatment. ↩