CAR-T's Boxed Warning for T-Cell Cancers: What the Evidence Does and Does Not Show

A boxed warning is the strongest caution the FDA can print on a drug label, and in early 2024 the agency placed one on all six approved CAR-T cell products (Kymriah, Yescarta, Tecartus, Breyanzi, Abecma, and Carvykti) for the risk of new T-cell cancers. If you read it carefully, that warning says something narrower than the headlines suggested. It records an association the agency judged serious enough to disclose and monitor. It does not prove that CAR-T causes these cancers, and it does not reverse the conclusion that, for the approved uses, the benefits still outweigh the known harms.

Key points#

• The boxed warning is a monitoring mandate, not a recall or a safety verdict.
• It rested on 22 reported T-cell cancers among CAR-T recipients as of the end of 2023, with three carrying the engineered gene inside the tumor.
• Direct molecular evidence exists in a handful of cases; a general cause-and-effect claim does not.
• Prior chemotherapy, the underlying blood cancer, and longer survival all compete to explain the same signal.
• Against tens of thousands of people treated, second T-cell cancers remain uncommon.

What a boxed warning is actually claiming#

It helps to separate two things a label can do. One is to describe a hazard that has been established. The other is to require disclosure of a risk that regulators consider plausible and serious while the science is still resolving. The CAR-T boxed warning belongs to the second group. It obliges manufacturers and clinicians to counsel patients and to monitor them for life, on the reasoning that a rare but grave outcome deserves prominent notice even before its frequency or mechanism is pinned down. If you treat the warning as proof of causation, you are reading more into the label than the FDA put there.

How a signal became a class-wide action#

The timeline shows a regulator escalating in steps. Late in 2023 the FDA said it was looking into reports of T-cell cancers, some of them CAR-positive, following treatment with CD19-directed and BCMA-directed autologous CAR-T therapies. Within a couple of months it had issued labeling-change letters requiring the warning across the whole class. Officials from the agency's biologics center later laid out the underlying tally in a medical-journal perspective: 22 cases known by the close of 2023, gathered from adverse-event reports and published papers.

Of those, fourteen had enough detail for closer study. Every one of the fourteen appeared within two years of the infusion, and about half within the first year. That tight clustering in time is the feature that carried the signal over the line. Applying the label's most prominent caution to a class of therapies that has changed the outlook for relapsed leukemias, lymphomas, and myeloma was not a reflex decision.

The three cases that carried the most weight#

The smallest figure in the whole account is also the most striking. In three reported cases, sequencing found the CAR transgene inside the malignant T-cell clone itself. That is a mechanistic clue worth taking seriously. Making a CAR-T product means using a viral vector to splice the receptor gene into a patient's own T cells, and any integrating vector carries a theoretical risk of insertional mutagenesis, meaning the new gene lands next to a growth-driving gene and helps switch it on. Finding the transgene woven into the tumor is consistent with the product having a hand in that particular cancer.

Consistent with, though, is not the same as responsible for. Two case reports published within months of each other reached opposite verdicts. One found no CAR gene in the lymphoma and judged the therapy blameless for that tumor. The other found the CAR gene together with pre-existing mutations and judged that the treatment probably contributed. Both can be right, because they describe different patients and different cancers. The fair summary is that insertional mutagenesis looks real in a small subset and unproven as a general rule.

The limits of counting reports#

Much of the public discussion draws on disproportionality analysis of spontaneous-reporting databases. A 2024 study pooled second-cancer reports from the FDA's FAERS system and the WHO's VigiBase, pulling 310 cases from the former and 297 from the latter across 2017 to 2023. Among the second cancers examined, T-cell lymphoma stood out with the strongest reporting odds ratio, with fainter signals for myelodysplastic syndrome and acute myeloid leukemia.

A reporting odds ratio asks whether an event shows up more often than expected for one drug compared with every other drug in the database. It is a tool for deciding where you should look, not a measure of how often something happens. These systems carry reporting bias, have no count of everyone actually treated, and frequently miss full treatment histories. That last gap is the decisive one. Nearly every CAR-T recipient has already been through lymphodepleting and cytotoxic chemotherapy, which on its own lifts the long-term odds of a second cancer, and most carry blood cancers that have accumulated their own mutations. A high odds ratio points a finger; it cannot tell the vector's contribution apart from the chemotherapy, the original disease, or the plain fact that people who survive longer have more time to develop something new.

Keeping the numbers in proportion#

Scale restores perspective. More than 34,000 people have now received CAR-T therapy, and the National Cancer Institute has stressed that second cancers of any kind stay uncommon in this group. One large single-center series followed more than 700 treated patients across nine years and counted 25 second cancers in all, only one of them a T-cell lymphoma. Set against tens of thousands of recipients, 22 reported T-cell cancers describe a low-frequency event, and low-frequency events are exactly the setting where causation is hardest to establish and where a cautious label earns its place.

What the evidence supports so far#

Three things stand on solid ground: the clustering of cases in the first two years is real, a class-wide warning is a defensible answer to that clustered signal, and at least a few cases carry molecular fingerprints of product involvement. Three things do not: general causation is unproven, disproportionality signals cannot yield a true rate, and the share owed to earlier chemotherapy and underlying disease cannot yet be separated out. Keep those two lists apart when you explain this to a patient. What will eventually turn this signal into a number is not a pile of spontaneous reports but prospective registries with complete treatment histories and long follow-up.