In-Vivo CAR-T Cell Therapy: How Engineering T-Cells Inside the Body Works

This article is science communication for general and technical readers. It is not medical advice; treatment decisions belong with a qualified clinician.

In-vivo CAR-T cell therapy flips the most expensive idea in modern oncology on its head. Instead of extracting a patient’s T cells, shipping them to a factory, rewiring them with a chimeric antigen receptor gene, growing them for a week or two, and infusing them back, the in-vivo approach injects a targeted drug that reprograms those same T cells while they are still circulating in the bloodstream. No apheresis machine. No clean-room manufacturing slot. No two-to-four-week vein-to-vein wait during which a patient’s disease keeps progressing. The therapy becomes something closer to a vial you pull off a shelf. If it works at scale, it could turn a bespoke $400,000-plus procedure available at a few dozen certified centers into an off-the-shelf injection — while raising a fresh set of hard questions about targeting, safety, and how long an engineered cell needs to live to cure someone.

What this covers: what a CAR is at the molecular level, why ex-vivo manufacturing is the real bottleneck, how lipid nanoparticles and targeted viral vectors engineer T cells in situ, the transient-versus-permanent trade-off, the safety failure modes, and how to read the 2026 pipeline without getting spun.

Context and Background

Chimeric antigen receptor T-cell therapy is one of the genuine triumphs of translational immunology. Beginning with the first FDA approval in 2017, CAR-T products have driven deep, durable remissions in patients with relapsed or refractory B-cell leukemias and lymphomas — cancers that had exhausted every other option. The approach later extended to multiple myeloma using receptors aimed at the BCMA antigen. For a subset of these patients, a single infusion of engineered cells did what chemotherapy, radiation, and stem-cell transplant could not. That is not incremental; it is a categorical shift in what a “living drug” can accomplish.

The problem is that the drug is literally alive, and manufacturing living cells for one patient at a time is brutally hard. Each dose is autologous — made from that specific patient’s own T cells. The workflow demands leukapheresis to harvest cells, cryopreservation, shipment to a centralized facility, genetic modification, days of expansion in bioreactors, sterility and potency release testing, shipment back, and a course of lymphodepleting chemotherapy before the cells go in. The whole loop is the “vein-to-vein time,” and it commonly runs two to four weeks. During that window, aggressive disease does not politely wait.

The economics compound the biology. List prices sit in the mid-six figures before hospitalization and toxicity management. Certified treatment centers are concentrated in wealthy urban hubs. Manufacturing capacity is finite, so slots are rationed. A meaningful fraction of patients deteriorate or die before their cells are ready — the sobering term of art is “drop-out,” patients who never receive the product they were harvested for. There is also a quality problem hiding in the logistics: heavily pre-treated patients often have exhausted, dysfunctional T cells, so the raw material for their own therapy is compromised before manufacturing even begins. For a therapy this effective, the binding constraint is not efficacy — it is access, cost, time, and starting-material quality. That is precisely the constraint the in-vivo strategy attacks, and it is worth noting that in-vivo engineering does not automatically fix the exhausted-cell problem; it simply removes the manufacturing bottleneck around it. The same forces are reshaping adjacent fields; see how cell-free biomanufacturing is rethinking protein production by removing living cells from the loop entirely. For the clinical foundation of CAR-T, the National Cancer Institute’s CAR T-cell overview is a solid primer.

How In-Vivo CAR-T Works

In-vivo CAR-T cell therapy delivers the CAR-encoding genetic payload directly into the patient’s body using a targeted carrier — most often a lipid nanoparticle loaded with mRNA, or a viral vector coated with a T-cell-seeking ligand. The carrier finds T cells in the bloodstream, hands off the genetic instructions, and the cell begins expressing the chimeric antigen receptor on its own surface. The factory step disappears because the patient’s body becomes the reactor.

Figure 1: The modular anatomy of a chimeric antigen receptor, from the antigen-binding scFv at the top to the CD3zeta activation domain inside the cell.

Figure 1 walks the CAR from outside the cell to inside it. Each module is a distinct engineering choice, and the same construct is used whether the cell is engineered in a factory or in a vein — the delivery route changes, not the receptor design.

The CAR construct, module by module

A chimeric antigen receptor is a synthetic protein that fuses the recognition ability of an antibody with the killing machinery of a T cell. It has four functional regions arranged as a single chain.

At the tip sits the single-chain variable fragment (scFv) — the variable heavy and light chains of an antibody stitched together. This is the part that physically grabs the target antigen, such as CD19 on B cells or BCMA on plasma cells. Its specificity determines what the cell will attack, so scFv choice is the single most consequential design decision in the whole receptor.

Below the scFv is the hinge, a flexible extracellular spacer. It sets the distance and freedom of motion the binder needs to reach antigens sitting at different heights on a target cell, and it influences how tightly the immune synapse forms. Next is the transmembrane domain, which anchors the receptor in the cell membrane and, in practice, affects receptor stability and clustering.

Inside the cell live the signaling domains. The costimulatory domain — usually derived from 4-1BB (CD137) or CD28 — provides the “second signal” that shapes how the cell responds. CD28 tends to drive fast, intense activation and rapid killing; 4-1BB tends to favor slower onset but greater persistence and a more memory-like phenotype. Finally, the CD3zeta domain delivers the primary activation signal that tells the T cell to proliferate and destroy whatever the scFv has bound. This four-part logic — bind, space, anchor, activate — is what makes a CAR a programmable kill switch rather than a generic immune stimulant.

Engineering the cell in situ instead of in a factory

The insight behind in-vivo CAR-T is that the genetic instructions for that receptor do not need to be installed in a clean room. If you can deliver the CAR-coding sequence selectively to T cells inside the body, the cells will build the receptor themselves. This collapses apheresis, transduction, expansion, and quality release into a single administration event.

Two hard problems make this non-trivial. First, selectivity: the carrier must reach T cells and largely spare everything else, because engineering the wrong cell type is both a waste and a safety hazard. Second, payload behavior: whether the CAR is expressed briefly or permanently depends on whether the payload is mRNA or an integrating DNA vector, and that choice cascades into durability and risk. The remainder of this section covers the delivery vehicles; the next section covers the payload trade-off in depth.

Why is this suddenly plausible in the 2020s when the idea is decades old? Three enabling technologies converged. The first is mRNA formulation and stabilization — the same chemistry, including modified nucleosides and optimized cap structures, that made mRNA vaccines work also makes a CAR-encoding message stable enough to survive delivery and translate efficiently. The second is ionizable lipid design: the ability to build LNPs that stay neutral in the bloodstream but become charged inside the acidic endosome, which is what drives the endosomal escape step in Figure 3 and lets the payload actually reach the cytosol instead of being digested. The third is antibody-fragment engineering, which supplies the small, stable binders that can be grafted onto a nanoparticle surface or a viral envelope to redirect it toward T cells. None of these is unique to CAR-T, but together they turned “engineer a specific cell type in situ” from a thought experiment into a laboratory reality.

LNP-mRNA versus targeted viral vectors

There are two dominant delivery families, and they sit at opposite ends of a durability spectrum.

Lipid nanoparticles carrying mRNA are the same broad technology class that powered mRNA vaccines. An LNP is a tiny fat sphere built from ionizable lipids, helper lipids, cholesterol, and a PEG-lipid, packaging a strand of messenger RNA that encodes the CAR. To make it T-cell-selective, researchers decorate the surface with a targeting ligand — for example an antibody fragment against CD5, CD7, CD3, or CD8 — so the particle preferentially docks onto T cells. Because mRNA never enters the nucleus and never integrates into the genome, the CAR it produces is transient: the cell translates the message for a few days, then the mRNA degrades and CAR expression fades. That is a feature for safety and a challenge for durability.

Targeted viral vectors — typically lentivirus, or in some designs adeno-associated virus (AAV) — take the opposite tack. A lentiviral vector integrates the CAR gene into the T cell’s genome, so once a cell is transduced it and its daughter cells express the CAR permanently. To make an ordinarily promiscuous virus T-cell-specific, the vector’s envelope is re-engineered (“pseudotyped”) to display a T-cell-targeting binder, such as an anti-CD3 or anti-CD8 domain, in place of its natural tropism. Get the retargeting right and you can, in principle, generate long-lived CAR-T cells with a single injection — no ex-vivo step at all.

There is a subtlety that separates the two families beyond durability: resting T cells are hard to transduce. In the ex-vivo workflow, cells are deliberately activated and pushed to divide before a lentivirus is added, because integration is far more efficient in dividing cells. An in-vivo vector has no such luxury — the T cells it meets in the bloodstream are mostly quiescent. Some designs address this by pairing the vector with a signal that transiently activates the cell as it is transduced, so the engineering and the priming happen together. This is one reason AAV, which does not require cell division to express its payload but also does not efficiently integrate, occupies an awkward middle ground: it can drive expression in resting cells but tends to be lost as those cells divide, giving a semi-durable window rather than a permanent one. The delivery vehicle, in other words, is not just a courier — its biology shapes which T cells get engineered and for how long.

A third, less-discussed vehicle deserves mention: cell-targeted LNPs delivering not mRNA but a gene-editing payload. Instead of shipping a CAR message, some experimental designs ship the components to knock in a CAR at a defined genomic site or to knock out an inhibitory gene, blending in-vivo delivery with site-specific editing. These sit even earlier on the maturity curve, but they hint at where the field is heading: precise, programmable, in-place cell engineering rather than crude gene insertion.

Deeper Analysis: Delivery, Targeting, and Safety

The transient-versus-permanent choice is the fault line that runs through every design decision in this field, so it is worth pulling apart carefully before looking at how carriers actually find their targets and where safety goes wrong.

Figure 3: A targeted lipid nanoparticle binds a T-cell surface marker, is taken up by endocytosis, escapes the endosome, and its mRNA is translated into a transient CAR that fades as the message degrades.

Figure 3 traces the LNP-mRNA route step by step. The particle’s targeting ligand binds a T-cell marker; the cell internalizes it by endocytosis; the ionizable lipids help the payload escape the endosome into the cytosol; ribosomes translate the mRNA into CAR protein; and within days the mRNA degrades, so the therapeutic window is inherently self-limiting.

Transient mRNA versus an integrating vector

An mRNA CAR is loud but short. Expression peaks within a day or two and is essentially gone within a week, because the cell dilutes and degrades the message and cannot copy it. That transience is a genuine safety asset. If a patient develops a severe toxicity, you do not have a self-replicating population of engineered cells that will persist for years — you wait, the CAR clears, and the T cell returns to baseline. Because nothing integrates into the genome, the theoretical risk of insertional mutagenesis (a vector landing inside a growth-controlling gene and nudging a cell toward malignancy) is removed. The cost is durability: a single transient dose may not sustain the weeks-to-months of continuous pressure needed to clear a large tumor burden, which is why transient designs are often framed around repeat dosing.

An integrating vector is the mirror image. One successful transduction can yield a self-renewing population of CAR-T cells that persist and patrol for months or years — the property most associated with durable cures in the ex-vivo world. The price is permanence in every sense: you cannot switch it off, insertional mutagenesis is a non-zero risk that regulators scrutinize closely, and any manufacturing or targeting error is baked into the patient’s genome. The mechanistic tension is clean — mRNA trades durability for reversibility and genomic safety; integrating vectors trade reversibility and that safety margin for durability.

It is worth being precise about why transient expression fights against efficacy, because the reason is quantitative, not merely temporal. Clearing cancer is a race between two exponential processes: the rate at which CAR-T cells find and kill tumor cells, and the rate at which the tumor grows and adapts. In the ex-vivo world, infused cells not only persist but often expand in vivo after infusion, multiplying by orders of magnitude as they engage antigen — that expansion is a large part of why a single dose can be curative. A pure mRNA CAR cannot expand its engineered receptor pool the same way, because dividing cells dilute the non-replicating message between daughters. So even if you engineer many T cells at once, the total CAR “dose” trends downward from the moment of injection rather than upward. Designs that want durability from a transient payload therefore either re-dose to keep topping up the pool, or combine mRNA delivery with a signal that drives the engineered cells to proliferate before the message fades. Understanding this asymmetry is the key to reading any in-vivo efficacy claim: expression is a snapshot; cure is an integral over time.

There is also a middle path some programs pursue — a self-amplifying RNA (saRNA) payload. Self-amplifying constructs carry, alongside the CAR sequence, a replicase that copies the RNA inside the cytosol, extending expression from days toward weeks without ever entering the nucleus or integrating. In principle this narrows the durability gap while preserving the genomic-safety advantage of RNA. In practice, the replicase adds immunogenic protein and its own manufacturing complexity, and controlling how long the amplification runs is non-trivial. saRNA is a good example of the field refusing to accept the transient-versus-permanent choice as binary and engineering its way toward the middle.

Figure 2: The ex-vivo pathway (apheresis, transduction, expansion, release testing, lymphodepletion, infusion) versus the compressed in-vivo pathway (inject a targeted carrier, engineer cells in situ, express the CAR).

Figure 2 makes the logistical asymmetry visible: a multi-step, multi-week, per-patient manufacturing chain on one side, and a single injection on the other. The clinical prize of the right-hand path is not just cost — it is time, and time is survival in aggressive disease.

It is worth naming a second-order benefit that is easy to miss: the in-vivo path may partly sidestep lymphodepletion. In the ex-vivo workflow, patients receive a round of lymphodepleting chemotherapy before infusion to “make room” for the engineered cells and to reduce competing immune cells that would otherwise limit their expansion. That chemotherapy is itself a source of toxicity and delay. If in-vivo engineering can generate an effective CAR-T response without requiring aggressive lymphodepletion — because it works with the patient’s existing, resident T-cell pool rather than a re-infused product — it removes not just a manufacturing step but a whole conditioning regimen. Whether that holds in practice is an open empirical question, and it is one of the more consequential things to watch, because a therapy that skips both the factory and the conditioning chemotherapy would look radically simpler than anything available today.

Tropism and targeting specificity

Selectivity is the whole ballgame. A carrier that engineers hepatocytes or myeloid cells instead of T cells is at best inert and at worst dangerous. Two levers control this. The first is the targeting ligand — the antibody fragment on the LNP surface or the retargeted viral envelope — which biases binding toward cells displaying a chosen T-cell marker like CD5, CD7, CD3, or CD8. The second is intrinsic biodistribution: unmodified LNPs tend to accumulate in the liver, so overcoming that default tropism is an active engineering fight involving lipid composition, PEGylation, particle size, and surface charge.

No targeting is perfect. Some fraction of particles will reach non-target cells, so designers plan for off-target transduction rather than pretending it away. This is where the transient-versus-permanent choice re-enters: an off-target mRNA event is self-correcting within days, while an off-target integration is forever. Selectivity and payload reversibility are therefore complementary safety layers, not independent ones.

A useful way to think about targeting is that it operates in two stages, and both have to succeed. The first is binding specificity — does the ligand latch onto the intended marker and not onto lookalike receptors on other cells? The second is functional selectivity — even after binding, does the particle actually deliver a working payload into that cell, or does it get bound, sequestered, and degraded without productive uptake? A particle can be beautifully specific in a binding assay and still fail functionally, or bind loosely yet deliver efficiently. The metric that matters clinically is not “what fraction of particles touched a T cell” but “what fraction of functional CAR expression ended up in T cells versus everything else.” Programs that report only the former are telling you less than they seem to.

The choice of which T-cell marker to target is itself a design trade-off. Broad markers like CD3 hit essentially all T cells, maximizing the pool you can engineer but offering no discrimination between helpful and unhelpful T-cell subsets. Narrower markers like CD8 bias toward cytotoxic “killer” T cells, which is often what you want for tumor killing, but shrink the addressable pool and may miss the helper and memory subsets that support persistence. There is no universally correct answer; the right marker depends on the disease, the payload, and whether the goal is a fast cytotoxic burst or a durable, self-sustaining response.

Engineering in safety: logic gates and off switches

Because in-vivo CAR-T sacrifices the pre-infusion control point, designers try to build safety into the receptor itself. Several strategies carry over directly from the ex-vivo world and become even more valuable when you cannot characterize the dose beforehand.

Logic-gated CARs require two conditions before they fire. An AND gate splits recognition across two receptors — a low-affinity primary CAR plus a co-receptor — so the T cell only fully activates when both antigens are present, sharply narrowing on-target off-tumor damage to cells that display the specific antigen combination found on the tumor. A NOT gate does the inverse: an inhibitory receptor recognizing an antigen found on healthy tissue vetoes killing there, sparing normal cells that happen to share the primary target. These circuits trade simplicity for precision, and precision is exactly what a poorly characterized in-vivo dose needs.

Suicide switches are the blunt-instrument backstop. A gene like inducible caspase-9 can be co-delivered so that, if toxicity spirals, a small-molecule drug triggers the engineered cells to self-destruct. With transient mRNA this matters less — the CAR clears on its own — but with an integrating in-vivo vector, an inducible off switch is close to mandatory, because it is the only way to reverse a permanent modification you cannot physically remove. The recurring theme is that in-vivo delivery pushes more of the safety burden onto the molecular design of the construct, since the procedural safeguards of the factory model are gone.

Immune barriers, CRS, and on-target off-tumor risk

The body actively resists these carriers. Pre-existing anti-AAV antibodies can neutralize viral vectors in a large share of the population. LNP components — the PEG-lipid in particular — can provoke antibody responses that blunt or complicate repeat dosing, which is a real constraint precisely because transient mRNA designs often depend on repeat dosing. Any protein-based targeting ligand can itself be immunogenic.

Then there are the toxicities intrinsic to CAR-T regardless of how the cell was made. Cytokine release syndrome (CRS) is a systemic inflammatory storm triggered when large numbers of activated CAR-T cells release cytokines such as IL-6, IL-1, and interferon-gamma; it ranges from fever to hypotension, capillary leak, and life-threatening multi-organ dysfunction. Immune effector cell-associated neurotoxicity syndrome (ICANS) is a related complication involving confusion, aphasia, and in severe cases cerebral edema. And on-target off-tumor toxicity occurs when the target antigen is also present on healthy tissue — the CAR does exactly its job, but on the wrong cells. The canonical example is B-cell aplasia: CD19-directed CAR-T destroys healthy B cells along with the malignant ones, a tolerable trade-off for a curable leukemia but a cautionary tale for antigens on tissues you cannot afford to lose.

In-vivo delivery does not eliminate any of these; it changes the calculus in both directions. On the reassuring side, transient expression may cap the severity and duration of CRS, because there is no self-amplifying cell population to sustain the cytokine storm — the pressure eases as the CAR clears. On the worrying side, engineering cells directly in the bloodstream removes the ex-vivo dose-characterization step. In the factory model you count and phenotype your product before it goes in; in the in-vivo model you infer the dose from indirect readouts, because the “product” is assembled inside a living, variable patient. Two patients given the identical injection may end up with very different numbers of CAR-T cells depending on their T-cell counts, receptor density, and immune status. That variability is a genuine control problem, and it is why dose predictability, not just average efficacy, is a metric to watch closely.

Dimension	Ex-vivo CAR-T	In-vivo CAR-T
Manufacturing	Per-patient clean-room, days of expansion	Off-the-shelf carrier, no cell factory
Vein-to-vein time	Weeks	Hours to a single visit
Cost driver	Bespoke cell manufacturing	Carrier (LNP/vector) production
Durability	Often long-lived, integrating	mRNA transient; vector can persist
Reversibility	Limited once infused	High with mRNA; low with integration
Dose control	Characterized before infusion	Inferred; harder to measure in situ
Off-target risk	Contained to the manufactured product	Depends on carrier tropism in the body
Maturity (2026)	Approved, standard of care	Largely preclinical and early clinical

Trade-offs, Gotchas, and What Goes Wrong

The headline risk is durability of transient CARs. Curing bulky disease historically requires CAR-T cells to persist and keep killing for weeks. An mRNA CAR that vanishes in days may control disease only if re-dosed on a schedule, and every re-dose invites anti-carrier immunity that can progressively neutralize the therapy. Solving durability without giving up the reversibility that makes mRNA safe is the field’s central unsolved tension.

Off-target transduction is the next trap. Because engineering happens in a living body full of many cell types, a carrier that is 95% selective still hits a lot of wrong cells at therapeutic doses. With integrating vectors, an off-target hit is permanent; with mRNA it self-clears — which is a large part of why several groups favor mRNA for the earliest in-vivo work.

Manufacturing simplification is relative, not absolute. In-vivo CAR-T removes the per-patient cell factory but adds a demanding chemistry-manufacturing-and-controls burden: reproducibly producing targeted LNPs or retargeted viral vectors at clinical scale and purity is its own hard problem, and the field’s tolerance for variability is low. The complexity moves upstream rather than disappearing.

Finally, regulatory and safety scrutiny is intense and appropriate. Delivering gene-modifying agents systemically raises questions about biodistribution, germline exposure, insertional risk for integrating designs, and immunogenicity of repeat dosing. Most in-vivo CAR-T programs in 2026 are preclinical or in early clinical stages — the science is genuinely promising, but durable-efficacy and long-term-safety data in humans are still maturing. Treat any claim of equivalence to approved ex-vivo therapy as investigational until that data exists.

A related gotcha is immunogenicity that compounds over time. A single injection of a targeted LNP or vector may work well, but the body remembers. Anti-PEG antibodies, anti-capsid antibodies, and antibodies against the targeting ligand or the CAR itself can accumulate, so the second and third doses of a repeat-dosing regimen may be progressively neutralized before they reach a T cell. This creates a perverse dynamic where the very transience that makes mRNA safe forces repeat dosing, and repeat dosing erodes the delivery efficiency the therapy depends on. Solving it may require rotating carriers, immune-silent lipid chemistries, or transient immunosuppression around dosing — each of which adds its own complexity and risk. This is not a footnote; it may be the single hardest problem standing between an elegant preclinical result and a durable human therapy.

Figure 4: A targeting-and-safety decision flow — antigen choice, delivery tropism, on- and off-target risk, and the mitigations (transient dosing, logic-gated CARs, tighter ligands) that feed a preclinical review.

Practical Recommendations

If you are following in-vivo CAR-T as an investor, clinician, student, or curious reader, the discipline is separating mechanism from marketing. The science is real and the logic is sound, but the gap between an elegant delivery paper and a durable human cure is wide, and press releases rarely map that gap honestly.

Watch three things. First, durability data — does a single or repeat dose actually clear disease and keep it cleared, or only show transient CAR expression? Expression is necessary but nowhere near sufficient. Second, selectivity evidence — what fraction of the intended payload reaches T cells versus liver and other tissue, and how is off-target engineering measured and bounded? Third, the payload choice and its rationale — mRNA versus integrating vector tells you almost everything about the durability, reversibility, and safety profile a program is betting on.

When you read a claim, apply this checklist:

Is it in humans yet? Preclinical mouse or non-human-primate data is encouraging, not decisive. Confirm the trial stage.
Transient or integrating? This single fact frames the entire safety and durability story.
What is the target antigen, and where else is it expressed? On-target off-tumor risk lives here.
How is selectivity quantified? “T-cell-targeted” should come with numbers, not adjectives.
Durable response or just detectable CAR? Insist on the distinction.
What are the CRS and immunogenicity data? Especially for any repeat-dosing regimen.

Hold in-vivo and ex-vivo to the same evidentiary bar. The promise is transformational; the proof, in 2026, is still being assembled.

Frequently Asked Questions

What is the difference between ex-vivo and in-vivo CAR-T?

Ex-vivo CAR-T removes a patient’s T cells, engineers them with a CAR gene in a manufacturing facility, expands them, and infuses them back — a multi-week, per-patient process. In-vivo CAR-T engineers the same T cells inside the body by injecting a targeted carrier (a lipid nanoparticle with mRNA, or a retargeted viral vector) that delivers the CAR gene directly to circulating T cells. The core benefit is eliminating the cell factory and the weeks of vein-to-vein time.

How do lipid nanoparticles deliver a CAR to T cells?

A lipid nanoparticle is a tiny fat sphere packaging mRNA that encodes the CAR. Its surface is decorated with a targeting ligand — an antibody fragment against a T-cell marker like CD5 or CD8 — so it preferentially binds T cells. The cell internalizes the particle, the payload escapes the endosome into the cytosol, and ribosomes translate the mRNA into CAR protein. Because mRNA does not integrate into the genome, the CAR is transient and fades within days.

Is in-vivo CAR-T therapy approved and available in 2026?

As of 2026, in-vivo CAR-T is largely investigational — most programs sit in preclinical or early clinical stages. Approved CAR-T products remain the ex-vivo type. Several biotech companies and academic labs are actively developing in-vivo approaches using both LNP-mRNA and targeted viral vectors, but durable-efficacy and long-term-safety data in humans are still maturing. This article is informational, not medical advice.

Why is transient CAR expression both good and bad?

Transient mRNA CARs are safer in one clear way: if a serious toxicity emerges, the engineered receptor clears on its own within days, and nothing integrates into the genome to cause long-term risk. The downside is durability — clearing bulky cancer often needs sustained CAR-T activity over weeks. A dose that fades quickly may require repeat administration, which in turn raises the risk of immune responses against the delivery carrier that reduce effectiveness over time.

What safety risks are specific to in-vivo CAR-T?

Beyond the CRS and neurotoxicity common to all CAR-T, in-vivo delivery adds off-target transduction (engineering unintended cell types), harder-to-measure dose because cells are made in the body rather than characterized beforehand, and immunogenicity against the carrier — particularly relevant for repeat dosing. On-target off-tumor toxicity, where the CAR attacks healthy tissue that shares the target antigen, remains a concern regardless of how the cell is engineered.

Could in-vivo CAR-T make cell therapy cheaper?

Potentially, by removing the per-patient clean-room manufacturing that drives the six-figure cost of ex-vivo CAR-T. An off-the-shelf carrier could be produced in batches and stored, more like a conventional biologic. But the savings are not guaranteed: making targeted LNPs or retargeted viral vectors reproducibly at clinical scale is its own demanding, costly process, so the complexity shifts upstream rather than vanishing. Real-world pricing will depend on how that manufacturing matures.

In-Vivo CAR-T Cell Therapy: How Engineering T-Cells Inside the Body Works (2026)